# Scam AI — full site content for LLMs and AI agents

> Scam AI (Reality Inc.) is a deepfake detection and scam-prevention platform: multimodal detection (video, images, audio) via the unified Eva V1.6 Detection Model API, Halo (an on-device meeting toolkit for Windows: meeting notes, speech-to-text, and real-time deepfake detection), KYC/IDV, and document-forgery detection, a separate plan priced per document and page and arranged with the team (https://scam.ai/contact) — the detection API does not take documents. Dashboard: https://app.scam.ai. Compact map: https://scam.ai/llms.txt

Marketing and docs content follows in full. App surfaces are served on app.scam.ai and are private.

# Platform

## Eva V1.6 Detection Model — our detection engine
URL: https://scam.ai/platform/eva

The unified API behind Scam AI's detection results: video, images, and audio, scored in real time.

### Trained on real attacks

Eva V1.6 learns from the largest library of real attack media in the industry — not synthetic benchmarks. When a new generation technique appears, the model retrains against it, so detection evolves at the speed of the threat.

### One API, every channel

Attacks rarely fake just one channel. The same Eva V1.6 call takes a photo, a video or a voice recording and routes it to its specialist detectors, so one integration covers swapped faces, generated images and cloned voices.

### Explainable results

Every score ships with the signals that fired. No black-box risk numbers — evidence your analysts and auditors can read and act on.

### Deployed where the risk is

The same engine runs two ways: as the unified API for batch and real-time flows, and on-device inside Halo for live meetings. One API, one set of results, wherever an attack can land.

### Inside a result: how Eva V1.6 scores media

When media arrives — a selfie, a video clip, a voice recording — Eva V1.6 breaks it into the surfaces where manipulation leaves traces: generation artifacts, blending boundaries where a swapped face meets the original frame, frame-to-frame inconsistencies no real camera produces. Each detector feeds one probabilistic score, and the response lists the signals that fired — an argument, not an assertion.

The exchange is simple: submit the file through the unified REST API and receive a JSON result with the score, the signal evidence, and per-frame detail for video. The same engine runs three ways — real-time, batch, and on-device inside Halo — so one API produces consistent results wherever an attack lands.

### Evaluating a detection model before you buy

Vendor demos are the worst place to judge a detection model — every vendor demos the fakes they already catch. Test Eva V1.6 the way you would test any risk model: assemble known-genuine media from real customers plus whatever confirmed fraud you have retained, and score it blind.

Then look past the headline result. Does every result carry evidence your analysts can read? Can you move the threshold yourself? Is retraining continuous as new face-swap and diffusion tools ship? Teams that buy well run shadow mode first, then step-up routing, then automated decisions as the score earns trust.

### FAQ

Q: What is the Eva V1.6 Detection Model?
A: The Eva V1.6 Detection Model is the detection engine behind every Scam AI result. Through one API it analyzes images, video, and voice recordings for signs of manipulation — face swaps, fully AI-generated media, and cloned voices — routing each to a specialist detector and returning a probabilistic score with the specific signals that fired. The model is retrained as new generation tools appear, and the same engine runs behind the cloud API, in on-prem deployments, and on-device inside Halo for live meetings.

Q: How does the Eva V1.6 Detection Model detect deepfakes?
A: The Eva V1.6 Detection Model looks for the traces synthetic media cannot help leaving: statistical artifacts from generation models, blending boundaries where a swapped face meets the original frame, temporal inconsistencies across video frames, and capture characteristics that do not match a physical camera. No single signal decides the result — the model weighs them together into one score and reports which ones fired. Because detection is probabilistic, you choose the threshold that fits your risk tolerance rather than accepting a fixed yes/no answer.

Q: Does the Eva V1.6 Detection Model give a yes/no answer or a score?
A: A score, with evidence. Every result includes a probabilistic score plus the list of signals that fired, so you can see why media was flagged, not just that it was. Your team sets thresholds on that score — auto-decline above one line, route to human review above another, pass below both — which keeps the false-positive and false-negative trade-off in your hands. For audit and compliance teams, the attached signal evidence makes each decision defensible after the fact.

## Check if AI — spot AI-generated media in seconds
URL: https://scam.ai/platform/check-if-ai

Upload an image and get an instant read on whether it's AI-generated.

### What does Check if AI do?

A fast, self-serve check for synthetic images: upload a JPEG, PNG, or WebP and Eva V1.6 returns a manipulation score with the signals behind it — no integration required. Video and audio checks run through the same unified API, not this page; document forgery is a separate plan.

### Who is Check if AI for?

A first-line gut check for trust & safety teams, journalists, and anyone who needs to know whether what they're looking at is real before they act on it.

### What happens when you hit upload

Behind the upload button is the same Eva V1.6 unified API that powers Scam AI's production integrations. Your image is read for the traces synthetic media leaves behind — generation artifacts, blending boundaries around swapped faces, and the edit signatures manual retouching leaves in pixel statistics. Seconds later you get a manipulation score with the signals that fired: the same result structure an enterprise integration receives.

This page takes images only, so anyone can test the engine without an integration; video and audio go through the same API. Your upload is saved to your account with the result, so you can open it again later.

### Reading a manipulation score like an analyst

A detection score is a probability, so think in likelihoods. A high score with multiple fired signals is strong evidence of manipulation; a low score with none is strong evidence of authenticity; a middling score is an honest uncertain — often what compressed, re-encoded, or low-resolution media produces, since each re-share strips the forensic detail detectors read.

The practical habits: check the signals, not just the number. Prefer the earliest, highest-quality copy of a file over a screenshot of a screenshot. Weigh the score alongside provenance — where the file came from and who benefits if it is believed. When the question becomes every upload, every day, move to the API.

### FAQ

Q: How can I check if an image is AI-generated?
A: Upload it to Check if AI. The file is analyzed by Eva V1.6, Scam AI's detection engine, which reads it for generation artifacts, blending boundaries, and manipulation signatures, and returns a manipulation score in seconds along with the specific signals that fired. No integration or setup is required. For the most reliable read, submit the earliest and highest-quality copy of the image you can find — screenshots and heavily re-shared copies lose the forensic detail that detection relies on.

Q: How reliable are AI-image detectors?
A: Honest ones report probability, not certainty. Detection reads statistical traces that generation and editing leave behind, and their strength varies with the file: an original-quality image gives the model much more to work with than a compressed, re-encoded copy. That is why Check if AI returns a score with named signals — a high score with several fired signals is strong evidence, a middling score is genuine uncertainty, and either should be weighed against where the file came from.

## Deepfake Playground
URL: https://scam.ai/platform/deepfake-playground

Deepfake yourself in 30 seconds. Point your camera at your own face, pick someone to become, and watch the swap happen live — the same thing an attacker can do with consumer tools.

### How easy it's become

A convincing face swap takes minutes — with free, consumer-grade tools. The playground lets you see that firsthand, so the threat stops being abstract.

### It's your own face, in real time

This isn't a canned clip. Your camera streams to a GPU running the swap frame by frame, and the result comes back live — you can move, talk, and change identity mid-session. That immediacy is the point: a threat you have personally produced stops being theoretical.

### FAQ

Q: How easy is it to make a deepfake?
A: Alarmingly easy — a convincing face swap takes seconds with consumer-grade tools and no technical skill. That accessibility is what changed the fraud landscape: deepfakes stopped being a specialist capability and became something any motivated attacker can produce at scale. The Deepfake Playground lets you experience it firsthand rather than take our word for it — you point a webcam at your own face and watch yourself become someone else in real time.

Q: Does the playground store my face or record the session?
A: No. Your camera streams to the swap engine for the length of the session so it can render each frame, and nothing is written down — no recording, no uploaded photo, no stored face. If you capture a shareable card, it is composed in your own browser from a frame already on your screen and never leaves your device unless you choose to share it.

Q: Can I test deepfake detection before buying?
A: Yes, in stages. The playground shows you the attack; Check if AI shows you the defence, letting you upload media for a scored read from the Eva V1.6 Detection Model — the same engine behind the production API. For a serious evaluation, run a proof of concept through the API with media from your own traffic, both known-genuine submissions and confirmed fraud, because your real distribution is what production performance depends on.

Q: Can the playground be used for security awareness training?
A: Yes, and it is one of its most common uses. Employees who have personally watched a convincing fake of themselves appear develop sharper instincts than any slide deck produces — especially finance, support, and executive-facing teams that fraudsters target with impersonation. Teams leave with one durable takeaway: seeing a familiar face on a call is no longer evidence that the person is who they appear to be.

## On-prem deepfake detection
URL: https://scam.ai/platform/on-prem-deployment

Run Scam AI detection inside your own environment — VPC, private cloud, or fully on-premises — so regulated data never leaves your perimeter.

### Detection where your data lives

For banks, government, and other regulated teams, media can't leave the building. On-prem deployment runs the same Eva V1.6 detection engine and API inside your environment, so results happen next to your data with no egress.

### Built for compliance

Meet data-residency, sovereignty, and procurement requirements with a self-hosted install: your keys, your network, your audit trail. Halo already runs on-device; on-prem extends that privacy posture to the full platform.

### Same platform, your perimeter

You get the same models, results, and integration surface as the cloud API, deployed to your infrastructure, updated on your schedule, and supported by our team.

### What a self-hosted install actually looks like

On-prem is the same Eva V1.6 engine and REST API as the cloud, packaged as containers that run in your VPC or air-gapped network, so media never leaves your perimeter and a cloud integration ports with no code changes. The one difference is model updates, which you apply on your own schedule.

### Due diligence for self-hosted detection

Judge self-hosted detection on how it handles isolation versus freshness: ask how often model updates ship, how they reach your environment, and what each upgrade asks of your team. Confirm the on-prem API is identical to the cloud one, and map the offering to your compliance requirements in writing.

### FAQ

Q: Can deepfake detection run fully on-premises?
A: Yes. Scam AI packages the same Eva V1.6 detection engine and REST API that power the cloud service for deployment inside your own environment — a VPC, a private cloud, or a fully air-gapped network. Your applications call a local endpoint with the same request and response shapes as the cloud API, and submitted media is scored entirely within your perimeter, with no egress. Results, evidence signals, and audit logs stay in infrastructure you operate, which satisfies data-residency and no-third-party-sharing requirements.

Q: How do detection model updates work in air-gapped environments?
A: Through packaged releases delivered over your approved transfer channels, applied on your schedule under your change-management process. This matters more for detection than for ordinary software: coverage depends on retraining against new deepfake and forgery generators, so a model frozen at install time gradually loses ground. When evaluating any self-hosted detection product, ask about release cadence, delivery mechanics for isolated networks, and how much work each upgrade demands from your operations team.

Q: When should we choose on-prem instead of the cloud API?
A: When data cannot leave your perimeter — data-residency laws, sovereignty mandates, sector regulation, or internal no-egress policy for sensitive media. If none of those apply, the cloud API is faster to adopt and carries no operational burden. A common middle path is hybrid: prove detection value with a cloud pilot on non-restricted traffic, then move regulated workloads on-prem once procurement and security review complete. Because the API surface is identical, migration does not require rebuilding the integration.

## PII Zero: Claims Media De-identification
URL: https://scam.ai/platform/pii-zero

Offline redaction of policyholder identity in claim documents, loss photos and video: names, addresses, account and policy numbers, faces and plates.

### Detect, track, redact, verify

OCR reads every page or frame in three orientations; NER, checksum recognisers and a vision layer flag names, IDs, faces, plates and codes. Regions are tracked across frames and blurred beyond recovery. The output is re-scanned with the same detectors and any surviving hit blocks the release.

### Measured on real footage

The product page shows actual output on synthetic claim documents and stock footage, with the number of tracked regions, frames covered and residual PII found by the second pass.

### Your perimeter, your audit log

Ships as containers for a VPC, private cloud or air-gapped network. Originals are read-only; the audit log records entity types, coordinates and a masked hint, never the full value. An optional encrypted vault keeps originals for a review window you set.

### FAQ

Q: Is the blur reversible?
A: No. Each region is downsampled to a few pixels before it is blurred, so the shapes are discarded rather than smeared, and the output is re-read by the same detectors to prove it. Opaque fill is available by policy.

Q: Does any data leave our environment?
A: No. OCR, NER, face and plate detection run as local models with no API calls. Nothing is retained after the run except the audit log you configure.

## Risk Signals
URL: https://scam.ai/platform/risk-signals

Turn detection results into a view of your threat surface — patterns rather than single incidents. Available on request.

### See where attacks concentrate

Aggregated across your checks, the same evidence shows which channels, teams, and workflows attract the most synthetic-media activity — so you can put defenses where attackers actually go. Set up with our team rather than self-serve.

### Trends, not noise

Attack techniques change weekly. The same result stream can be tracked over time to show how the threat mix shifts — new generation tools, new targets, new hours of attack.

### Close the loop

Signals are worth more fed back than filed. Use them to tune your own thresholds and route the riskiest sessions to review, so what you learn on Monday hardens your defenses by Tuesday.

### FAQ

Q: What are risk signals in fraud detection?
A: Risk signals are the structured evidence a detection system produces alongside each result — which manipulation indicators fired, on which channel, in which workflow, and when. Individually they explain one decision; aggregated, they describe your threat surface. Scam AI can roll your result telemetry up into heatmaps and trends — where synthetic-media activity concentrates, how the attack mix shifts week over week, and where the same forged template keeps reappearing. It is arranged with our team rather than switched on in the dashboard, which reports one verdict per check.

Q: How do teams monitor deepfake attack trends?
A: The raw material is result telemetry: every detection check records what fired, where, and when. Trend monitoring aggregates that stream by channel, team, and workflow, so shifts become visible — a new generation tool in onboarding, injection attempts concentrating on one product, attack volume moving to off-hours. Effective teams review trends on a fixed cadence with a named owner, then respond through policy: tightening thresholds on workflows under pressure or routing the riskiest sessions to human review.

Q: Can detection thresholds be tuned based on analytics?
A: Yes — that feedback loop is the point of trend data. Scam AI results are probabilistic scores rather than fixed yes/no answers, so thresholds are yours to set, and trends tell you when to move them. If trend data shows a campaign concentrating on one workflow, lower that workflow's auto-decline threshold or widen its human-review band — a policy change, not a redeploy. Best practice: decide in advance which patterns trigger which changes, so tuning is documented policy.

## Deepfake Detection
URL: https://scam.ai/platform/deepfake-detection

Purpose-built deepfake signals — face swaps, full generations, and replays — scored across images, video, and live calls.

Key stat: 1 / 5 min — a deepfake fraud attempt occurred every five minutes in 2024 (source: Entrust Identity Fraud Report)

### What it does

Eva V1.6 scores every selfie and video frame for synthetic-media signals, with coverage that tracks the latest diffusion and face-swap generators as they appear.

### Where it fits

IDV and onboarding stacks: score the selfie step before an account exists. The same signals run on claims footage, moderation queues, and live meetings through Halo.

### What you get back

A per-frame manipulation score with the signals that fired: face-swap, full generation, or replay. Your flow can auto-decline, step up, or route to a reviewer with the reason attached.

### What is deepfake detection?

Deepfake detection decides whether a face in an image, video, or live stream was synthetically generated or manipulated: a swapped face, an AI-generated person, or a replay passed off as live. Scam AI checks everywhere a face appears and returns a probabilistic score with the signals behind it.

### Choosing a detector: coverage, evidence, and rollout

A detector worth buying covers still images, recorded video, and live calls, not just the easy selfie, and names the signals behind each score so a decline is defensible. On rollout, score live traffic in shadow mode first and save automated declines for high-scoring results.

### FAQ

Q: Can deepfakes actually be detected?
A: Yes — with the honest caveat that detection is probabilistic, not absolute. Generation tools leave traces: statistical artifacts, blending boundaries, temporal inconsistencies in video, and capture characteristics no physical camera produces. A detector weighs those signals into a tiered likelihood result rather than a guaranteed one, and stays effective only if it retrains as new generators appear. Scam AI returns scores with evidence and lets you set thresholds — decline high-scoring fakes automatically, route ambiguous cases to human review.

Q: How does deepfake detection work on live video calls?
A: Recorded media can be scored after the fact; a live call has to be scored while it happens. Halo runs Scam AI's detection on-device during meetings, continuously checking participant faces and feeds for face swaps and manipulated video, and raising a risk signal the moment something synthetic appears — mid-call, while you can still act. Because analysis happens locally, the meeting never leaves the device to be checked, keeping sensitive calls private while they are protected.

Q: What kinds of deepfakes does detection need to cover?
A: Three broad families. Face swaps graft a target's face onto another person's body — the classic impersonation tool. Full generations create an entirely synthetic person, common in fake profiles and synthetic identities. Replays and injections present pre-recorded or synthetic footage as a live capture, often through virtual-camera software. Each leaves different evidence, so Scam AI scores them with distinct signals across images, recorded video, and live calls — catching the synthetic media itself, however it is delivered.

## ID Document Forgery Detection
URL: https://scam.ai/platform/id-document-forgery

Tampering and forgery signals for identity documents — passports, national IDs, driver's licenses, residence permits.

Key stat: 57% — of document fraud is now digital forgery rather than physical counterfeits (source: Entrust)

### What it does

Detects edits, template reuse, and full AI generation on government identity documents, returning scored signals with the evidence that fired.

### Evidence you can defend

Each flag returns the tampered region and the signal behind it, so a rejected document comes with a reason your compliance and audit teams can stand behind.

### The signals a forged ID leaves behind

A forged ID gives itself away in layers: a digital edit mismatches compression and lighting at its boundaries, a template forgery reuses the same artwork across IDs, and an AI-generated document carries model fingerprints. One API call reads all three and localizes the tampering, so a reviewer sees the evidence on the document itself.

### What to ask a document-forgery vendor

Document fraud has moved from physical counterfeits to digital edits and reused templates, so evaluate with your own corpus of real documents plus confirmed forgeries, in the types and countries you serve. Then check that a flag points at the tampered region, covers your document mix, and maps onto your approve, step-up, and decline logic.

### FAQ

Q: How does AI detect a fake ID?
A: By reading structure instead of content. An edited ID carries regions whose compression history, noise pattern, and lighting do not match the rest of the document. A template forgery reuses identical artwork across supposedly different IDs. A fully AI-generated document carries the statistical fingerprints of the model that produced it. Scam AI scores these signals together, returns a probabilistic score, and localizes the suspicious region — so a reviewer sees exactly which part of the document triggered the flag.

Q: Can AI-generated passports and driver's licenses be detected?
A: Yes. Fully generated documents look plausible at a glance but are structured like no issuing authority's real output — generation models leave statistical fingerprints, and synthetic documents often betray template inconsistencies a genuine document would never contain. Because new generation tools appear constantly, coverage depends on retraining: Scam AI's detection engine is updated as new document-generation techniques emerge. Results come back as scores with evidence rather than a blind pass/fail, so your team controls how aggressively to act.

Q: What should happen when a document is flagged in KYC?
A: Route by score, not by binary. High-scoring forgery results can auto-decline or trigger a hard stop; mid-band flags route to human review, where the localized evidence — the tampered region highlighted on the document — lets an analyst confirm or clear the flag in seconds. Every result and its signals should land in your audit trail, so a rejected applicant, a regulator, or an internal QA review can see the concrete reason behind the decision.

## Loan & Insurance Document Forgery
URL: https://scam.ai/platform/loan-insurance-forgery

Forgery signals for due-diligence documents — bank statements, pay stubs, proof of income and address, invoices, policy documents.

Key stat: $3.1B — U.S. lender exposure to synthetic-identity fraud in a single year (source: TransUnion)

### What it does

Scores the paperwork behind lending and claims decisions for tampering and AI generation, so doctored statements don't sail through underwriting.

### Where it fits

Loan origination, credit review, and insurance claims pipelines.

### Catch the pattern, not just the page

Because documents are scored across applications, Scam AI surfaces the same doctored template or synthetic income proof repeating across your book, exposing the organized fraud behind a single file.

### How a model reads a doctored bank statement

People check financial documents by whether the numbers look plausible, but the forgery lives in the structure: fonts and spacing that drift from the template, an edited figure whose noise doesn't match the page. Detection scores those structural signals with localized evidence, and across many files it surfaces the same template reused under different names.

### FAQ

Q: How can lenders detect forged bank statements?
A: Automated forgery detection reads the structure a human reviewer cannot see: fonts and alignment that drift from the bank's real template, edited regions whose noise and compression history differ from the rest of the page, and the statistical fingerprints of AI-generated documents. Scam AI scores each statement and returns a score with localized evidence — inside the origination flow via the REST API or in batch across a portfolio. Cross-application scoring exposes the same doctored template under different applicant names.

Q: Can AI detect fake pay stubs and proof-of-income documents?
A: Yes. Pay stubs, employment letters, and proof-of-address documents are edited or generated with the same tools as any other document, and they leave the same structural evidence: template inconsistencies, mismatched edit regions, and generation artifacts. Detection returns a probabilistic score with the signals that fired, localized to the suspicious area, so an underwriter reviews the flagged field rather than the whole file. Because results are threshold-based, you decide which scores auto-route to review and which pass without friction.

Q: How does document fraud detection fit into insurance claims processing?
A: At claims intake, every supporting document — invoices, repair estimates, policy paperwork, proof of loss — can be scored before adjudication begins. Genuine claims pass through without added friction; suspicious documents surface with a score and localized evidence attached, so adjusters spend their time on the files that need judgment. Batch endpoints also let teams re-screen historical claims, and cross-claim analysis flags documents and templates reused across supposedly unrelated claims — a common pattern in organized claims fraud.

## Age Estimation
URL: https://scam.ai/platform/age-estimation

Age checks from a live selfie — an estimated age and the probability the person is over your threshold, no ID collected.

### What it does

Estimates age from a live selfie with liveness checks, returning an estimated age and the probability the person is over your threshold, with no identity established and nothing stored.

### Where it fits

Age-gated signups, checkouts, and content access, verifying age without collecting an ID.

### Privacy by default

No document, no stored image, no identity established, just a live-selfie age estimate and an over-threshold probability you can gate on. Compliance stays light and data collection stays minimal.

### Estimating age from a face, not a database

Age estimation reads the facial features that change with age and maps them to an estimated age, plus the probability that the person clears your threshold, so it's estimation, not recognition: no database match, no identity, no biometric record. It runs on a live selfie in one API call, and the image is analyzed and discarded.

### FAQ

Q: How does AI age estimation work without documents?
A: The model reads visual features of a live selfie that correlate with age — skin texture, facial structure — and returns an estimated age together with the probability that the person is over your age threshold. No document is collected and no identity established: the system answers how old this person roughly is, not who they are. Scam AI pairs the estimate with liveness and capture checks so a photo, replay, or injected feed cannot substitute for a real face, and the selfie is analyzed and then discarded.

Q: Is facial age estimation the same as facial recognition?
A: No, and the difference matters for privacy and compliance. Facial recognition matches a face against a database to identify or verify a specific person, which creates and depends on biometric records. Age estimation infers one attribute — an approximate age — from the image in front of it, matches nothing, identifies no one, and stores no template. That is why estimation-based age gates are lighter for privacy teams to approve: a compliance signal without an identity system built around it.

Q: What happens when a user is close to the age threshold?
A: Estimation is probabilistic, so well-designed gates act on the over-threshold probability rather than a single cutoff on the estimated age. Users whose probability of clearing the required age is high pass automatically; when it is low, the gate holds. The uncertain middle routes to a stronger check — typically document verification — so the heavier, higher-friction step is reserved for cases that genuinely need it. In practice, most users clear the gate instantly from a selfie, and the ID upload becomes the exception path.

## Device Signals — liveness & presentation attacks
URL: https://scam.ai/platform/capture-signals

Detection for spoofed presentations — printed photos, screen replays, and masks — plus liveness failures, caught at the point of capture.

### What it does

Confirms the face at capture is a live, present person, not a printed photo, a screen replay, or a mask held to the lens. Spoofed presentations are flagged as the media is captured, before the frame enters your pipeline.

### Where it fits

The capture step of IDV: signup selfies, document photos, and liveness video, checked as the media is captured.

### The earliest line of defense

Running at capture means a spoofed or replayed presentation is caught before any downstream check even sees the frame, stopping the attack earlier, and cheaper, than catching it later.

### Layering liveness into an IDV stack

Liveness is one layer, not the whole defense, so pair it with deepfake and document detection: an attack that beats one check still has to beat the others. A failed check can trigger a retake or step-up instantly, because it's a fact about the capture, not a judgment about the person.

### FAQ

Q: What is a presentation attack?
A: A presentation attack presents a fake face to a real camera to pass as a live person: a printed photo, a video replayed on a screen, or a mask worn or held to the lens. It is the classic way naive selfie and liveness checks are defeated — the camera is real, but the thing in front of it is not a live subject. Liveness detection targets exactly this, reading the captured image for the signals a screen, print, or mask leaves behind.

Q: How does liveness detection work?
A: By examining the captured face rather than trusting the session. Detection reads the signals a spoof leaves — screen glare and moiré, print texture and edges, and the depth and micro-motion a living face shows that a photo or mask does not — and returns a scored signal alongside the media. Your flow can reject or step up a session that cannot demonstrate a live, present subject before the frames reach downstream analysis.

Q: How does this work with deepfake detection?
A: They are complementary layers. Liveness confirms a real, present human is in front of the camera; deepfake detection judges whether the face itself has been synthetically generated or swapped. Run together, an attack has to defeat both — a mask that survives a quick look still fails liveness, and a swapped face that looks live still fails deepfake detection. Scam AI returns each as its own scored signal with the result, so the decision comes with its reasons attached.

## Manual-Edit Forensics — image forgery detection
URL: https://scam.ai/platform/manual-edit-forensics

Pixel-level forensics for manually edited selfies and documents — splices, clones, and retouching.

### What it does

Not all forgery is AI: pixel-level analysis catches Photoshop-style edits like spliced regions, cloned patches, and retouching on selfies and documents.

### Where it fits

Alongside gen-AI detection, so hand-crafted fakes don't slip past a model looking only for synthetic artifacts.

### No gap to slip through

One pass covers both machine-made and hand-made forgery: Eva V1.6 scores synthetic generation while forensics catches manual edits, closing the seam attackers exploit between them.

### Splices, clones, and the physics of an edited image

Every manual edit disturbs an image's internal consistency: a splice imports noise and lighting that don't match, a clone leaves improbable repetition, and retouching breaks a camera's uniform sensor noise. Forensic analysis reads those pixel statistics to prove an edit happened and localize where, running in the same pass as Eva V1.6's AI detection.

### FAQ

Q: Can AI detect Photoshop edits?
A: Yes — through forensic signals rather than generative-AI fingerprints. Manual edits disturb an image's internal statistics: spliced content imports mismatched noise and compression history, cloned patches repeat pixels improbably, and retouching breaks the uniform sensor noise a real camera produces. Scam AI's forensics reads these traces on selfies and documents, returns a score with the signals that fired, and localizes the edited region — the flag points at the altered field, not vaguely at the whole image.

Q: What is image forgery localization?
A: Localization means identifying where in an image or document the manipulation occurred, not merely that it occurred somewhere. Instead of a file-level fake/real label, the result highlights the specific region whose forensic statistics are inconsistent — a pasted date field, a cloned background patch, a retouched face area. That precision changes operations: reviewers confirm or clear a flag in seconds, disputes can be answered with concrete evidence, and audit trails record exactly what was found rather than an opaque score.

Q: Why do I need edit forensics if I already detect deepfakes?
A: Because they catch different artifact classes. Deepfake detectors hunt the statistical fingerprints of generative models — and a hand-made edit contains none of them. A document field altered in an image editor can sail past a purely generative detector while remaining fully fraudulent. Attackers use whichever tool is cheapest for the job, so coverage needs both: Scam AI runs generative detection and manual-edit forensics in the same pass, giving a hand-made fake the same scrutiny as a synthetic one.

## MCP & Skills for LLMs
URL: https://scam.ai/platform/mcp

Give an AI agent one tool that checks whether an image, a video or an audio recording is real, and a Skill that tells it what to do with the answer.

### Connect in one step

Add the server to your agent's MCP configuration. Hosted, there is nothing to install: point the client at https://mcp.scam.ai/mcp and send your API key in an x-api-key header. On your own machine, run npx -y @scam-ai/mcp-server with the key in SCAMAI_API_KEY; it needs Node 22 and is the one that can read a local file path.

Both reach the same detection, use the same key and cost the same as a direct API call.

### One tool, and two free calls

detect_media checks one image, video or audio file, sent as a link, a local file path or the file's bytes, and works out which kind it is. get_credit_balance returns what the account has left, and get_usage lists recent credit movements; both are free.

A check is never retried for you: a retried check runs, and is billed, again.

### What comes back

Every answer carries a verdict to branch on. LIKELY_REAL means let it through, ALERT means put a human on it, and LIKELY_AI means block or escalate; audio answers LIKELY_REAL or LIKELY_AI only.

The score is the probability, from 0 to 1, that the media is AI-made or manipulated. It is not a confidence, and it is not proof. A one-sentence summary rides with it, and video adds a per-frame series and audio a per-window timeline, so a reviewer can see where in the clip the signal sits.

### Skills: what the agent does with the answer

The tool answers one question per file. A Skill is the short instruction file that tells the agent what to do with each answer: branch on the verdict, send anything that is not clearly real to a person, and never present a score as proof. Written into a Skill rather than a prompt, the rule holds however far the conversation drifts.

The Skill on this page is a starting point: copy it into your agent's skills folder and adapt its rules to your own review policy.

### What it costs

The same as a direct API call, billed in credits to the account behind the key, and a failed check that could not be judged is never charged. Document forgery detection is sold on its own plan.

### Where it fits

Anywhere an agent is about to rely on media that someone else supplied: the photos and recorded statement in an insurance claim, a selfie or ID video at onboarding or in a loan application, a candidate's interview recording, or a reported post in a trust and safety queue.

### FAQ

Q: What is an MCP server for deepfake detection?
A: MCP, the Model Context Protocol, is an open standard that lets an LLM or an agent call real tools. Scam AI's MCP server exposes one detection tool, detect_media: the agent passes an image, a video or an audio recording, as a link, a file path or the file's bytes, and gets back a verdict, a score from 0 to 1 and a one-sentence summary, with per-frame scores for video and a per-window timeline for audio. It answers exactly as the Scam AI API does and is billed the same way.

Q: How do I use a deepfake-detection MCP server in an AI agent?
A: Add the server to your agent's MCP configuration: the hosted URL, https://mcp.scam.ai/mcp, with your API key in an x-api-key header, or npx -y @scam-ai/mcp-server with the key in SCAMAI_API_KEY. Then give the agent one rule per verdict: continue on LIKELY_REAL, send ALERT to a person, and block or escalate LIKELY_AI. A Skill is the place to write that rule, so it holds however the conversation drifts.

Q: Which files can the Scam AI MCP server check?
A: Images (JPG, PNG or WEBP, up to 10 MB), video (MP4, MOV or WEBM, up to 100 MB) and audio (MP3, WAV, M4A, FLAC, OGG or AAC, up to 50 MB), or a link to a public video page or a media file. Document forgery detection, such as an edited PDF, is sold on its own plan and is not part of this tool.

Q: Which LLMs and agent frameworks work with Scam AI's MCP server?
A: Any client that speaks MCP, including Claude, Cursor and VS Code. The hosted server suits a client that cannot run a local process; the npx server runs on your own machine and can read a local file. For a platform with its own tool format, the same detection is one REST call, described by a public OpenAPI file at https://api.scam.ai/openapi.json.

## Detection API
URL: https://scam.ai/platform/api

One API for the places fraud happens: send an image, a video or a recording and get a scored verdict back. SDKs for Node and Python, a CLI and an MCP server are published.

### Simple to integrate

Submit media, receive a verdict, a score and the evidence behind it — per-frame for video, per-window for audio. Wire it into onboarding flows, support channels, claims pipelines, and content moderation.

### Built for production

Synchronous answers at real-time latency for live flows; for volume work call the same endpoint in parallel within your key's per-minute limit, and a repeated file within 30 days is never charged twice. Every result carries the evidence your risk team and your auditors can read.

### SDKs, CLI and MCP server

The REST API covers every language, and the official clients are published: @scam-ai/sdk for Node and TypeScript, scamai for Python, @scam-ai/cli for the terminal, and an MCP server — hosted or run locally — for AI agents.

Install lines, examples and the OpenAPI file are in the [docs](https://scam.ai/docs/sdks-and-dashboard).

### Running a proof of concept

Test on your own traffic — real genuine media plus known fraud — and score it blind. Then set the thresholds that decide what's auto-declined, reviewed, or approved.

### FAQ

Q: How do I integrate a deepfake detection API?
A: Send media over HTTPS and act on the JSON. Your backend submits an image, a video or a recording to the REST API with an API key and receives a verdict, a probability score and the evidence behind it — per-frame for video, per-window for audio. Your code compares the score to thresholds you set and branches — approve, step up, or route to review. There is no model to host; any language that can make an HTTP request can integrate today and the official Node and Python SDKs, the CLI and the MCP server are published.

Q: What does a deepfake detection API return?
A: A scored, explainable result rather than a bare label. The JSON response carries a probabilistic score for manipulation, a verdict to branch on, and per-frame detail for video and per-window detail for audio — the evidence behind the score — so you can locate where in a clip manipulation appears. That structure serves machines and humans alike: your code thresholds on the score for automated decisions, while analysts and auditors read the signals to defend each outcome after the fact.

Q: Does a deepfake detection API work in real time?
A: It should offer both speeds, because fraud workflows need both. Scam AI's synchronous endpoints return results at real-time latency for inline flows — an onboarding selfie check or a document step where a user is actively waiting. Volume work runs against the same synchronous endpoint in parallel, inside your key's per-minute limit; a file repeated within 30 days is not charged again. Both patterns return the same result structure, so one integration serves live decisions and pipeline jobs alike.

# Verifit

## Verifit | Remote inspections. Confident decisions.
URL: https://scam.ai/verifit

Cut inspection costs and keep decisions moving. Verify property and assets remotely, with fewer site visits and less back-and-forth.

### Less waiting between request and decision

Give customers an easy way to complete inspections and your team a faster way to review them.

### FAQ

Q: How does Verifit help reduce inspection costs?
A: Collect verified photos remotely to reduce travel, scheduling, and manual follow-ups. Your team gets what it needs to review suitable cases without an on-site visit.

Q: Does Verifit replace every inspection?
A: Use Verifit to handle suitable inspections remotely and reserve site visits for complex cases. Your team decides when more information or an in-person inspection is needed.

Q: How does my team get started?
A: Request a demo focused on your workflow. We will explore where Verifit can reduce delays, save your team time, and make inspections easier for your customers.

# Solutions

## KYC / IDV fraud detection
URL: https://scam.ai/solutions/kyc-idv

Make sure the person enrolling is live, present, and real — not a replay, a render, or a synthetic identity. Scam AI works with your KYC vendor to strengthen it, not replace it.

Key stat: $23B — projected annual U.S. losses to synthetic-identity fraud by 2030 (source: Deloitte Center for Financial Services)

### The threat

Face-swap kits and virtual cameras defeat naive selfie checks. Synthetic identities blend real and fabricated attributes to pass onboarding and mature into major fraud losses.

### Where detection sits in your onboarding flow

Scam AI plugs into the two moments identity fraud enters, the selfie capture and the document upload, calling the REST API with media you already collect and returning a scored result in real time. Clean scores flow through, borderline cases step up, and clear manipulation declines with evidence attached, so genuine applicants never see a new step.

### Breaking the synthetic identity lifecycle at day zero

A synthetic identity is grown, not stolen: fraudsters pair a real identifier with a fabricated face, behave normally for months, then bust out like a loyal customer defaulting. Enrollment is the only cheap place to stop it, because the fake face and doctored document must appear at day zero where each is detectable (Deloitte projects $23B in annual U.S. losses by 2030).

### FAQ

Q: How do fraudsters bypass selfie and liveness checks?
A: Attackers use virtual camera software to inject pre-recorded or AI-generated video into the verification session, bypassing the physical camera. Others use real-time face swaps that respond to liveness prompts, blinking and turning on cue, because a live person drives the fake face. Naive liveness checks confirm motion, not authenticity. Defeating these attacks means scoring the media itself — whether the face is synthetic and whether the presentation shows print, replay, or mask spoofing — so an injected or replayed fake is caught on its content. That deepfake and presentation-attack scoring is what Scam AI adds on top of standard liveness.

Q: What is synthetic identity fraud in KYC onboarding?
A: Synthetic identity fraud combines real identifiers, such as a government ID number, with fabricated names, faces, and documents to create a person who does not exist. Because part of the identity is real, database checks often pass, and the fraudster ages the account with normal activity before cashing out. Deloitte projects U.S. losses of $23B annually by 2030. At onboarding it depends on fake media — an AI-generated or swapped face and edited documents — the weak point Scam AI targets.

Q: Can a deepfake pass identity verification?
A: Consumer-grade face-swap tools defeat basic selfie comparison, and virtual cameras feed rendered video into verification flows that only check for a moving face. Deepfakes pass when the system trusts the video instead of inspecting it. They fail when verification scores the face for generation artifacts, checks the presentation for print, replay, or mask spoofing, and confirms consistency between selfie and document. Scam AI runs those checks at enrollment via the REST API, returning a scored result and its signals, so your team can decline or step up before the account exists.

## Loan & credit fraud detection
URL: https://scam.ai/solutions/loan-credit

Stop synthetic applicants and doctored income documents at origination, before the money moves.

Key stat: $3.1B — U.S. lender exposure to synthetic-identity fraud in a single year (source: TransUnion)

### The threat

Synthetic identities backed by forged pay stubs and doctored statements sail through underwriting. By the time repayment fails, the 'borrower' never existed.

### Plugged into origination, not bolted on after

Scam AI sits inside origination, before funds move: each pay stub, statement, or selfie is scored by the REST API with the signals behind it. Auto-decline clear forgeries, route borderline files to review, and pass clean applications untouched.

For guided remote inspections, explore [Verifit for lending](https://scam.ai/verifit/lending).

### Income documents are where the fraud lives

Most loan fraud needs fake income, not a fake person: a doctored payslip or inflated balance clears every identity check. Eva V1.6 reads each file for editing and AI-generation artifacts and flags one template reused across applications (TransUnion puts lender exposure at $3.1B a year).

### FAQ

Q: How do lenders detect fake pay stubs and bank statements?
A: Fake income documents are detected through forensic analysis rather than visual review: examining files for editing artifacts, font and layout inconsistencies, and signs of AI generation that persist even in convincing forgeries. Detection also compares documents across applications, since fraud rings reuse templates with only the numbers changed. Scam AI runs these checks through a REST API inside the origination flow, returning a per-document tamper result with evidence, so underwriters see which files need scrutiny instead of manually inspecting every PDF.

Q: What is synthetic identity fraud in lending?
A: In lending, synthetic identity fraud means a borrower who never existed: a blend of real and fabricated attributes, supported by forged pay stubs and statements, that passes underwriting, then defaults with no one to collect from. TransUnion measured U.S. lender exposure at $3.1B in a single year. Because the applicant cannot appear in person, the fraud depends entirely on media — a rendered face at verification and doctored documents — and both are detectable at origination, before the loan funds.

Q: Can loan application fraud be detected before funding?
A: Application fraud is most detectable at origination, because that is when the fabricated material is submitted at once: the selfie, the ID, and the income documents. Screening each for manipulation at upload catches the fraud while the decision can still change cheaply — a decline or a step-up rather than a charge-off. Scam AI returns results in real time inside the approval flow, and batch endpoints let lenders rescan existing books for synthetic applicants approved before detection was in place.

## Insurance claims fraud detection
URL: https://scam.ai/solutions/insurance-claims

Verify claim photos, videos, and documents are genuine — not staged, edited, or AI-fabricated.

Key stat: $308B — the annual cost of insurance fraud in the U.S. (source: Coalition Against Insurance Fraud)

### The threat

Claimants generate damage photos that never happened, edit real images to exaggerate a loss, and fabricate supporting paperwork with consumer AI tools.

### From FNOL to result without slowing the claim

Detection runs at first notice of loss: each uploaded photo, video, or document is scored by the REST API with evidence attached. Clean media stays straight-through, and flagged media routes to an adjuster or SIU without changing the claimant experience.

For guided remote inspections, explore [Verifit for insurance claims](https://scam.ai/verifit/insurance).

### The new fraud is editing reality, not inventing it

The hard problem is exaggeration, not fabrication: AI-extended damage or one changed digit passes every metadata check, because only the pixels lie. Eva V1.6 detects localized manipulation inside otherwise genuine images and localizes the edited region (insurance fraud already costs the U.S. $308B a year).

### FAQ

Q: How do insurers detect AI-generated or edited claim photos?
A: AI-generated and edited claim photos are caught by forensic analysis of the image itself: generation artifacts, inconsistencies in lighting and texture, signs of localized editing, and traces left by manipulation tools. Scam AI applies these checks to every photo, video, and document at upload through the REST API, returning a scored result with evidence attached, so adjusters know which files to trust before adjudication begins.

Q: What is a shallowfake in insurance fraud?
A: A shallowfake is a genuine photo or document altered with basic editing tools rather than generated from scratch — real damage extended, a date changed, an amount inflated. Shallowfakes are harder to catch with context checks because the underlying media is authentic; the manipulation is localized. Detecting them requires pixel-level forensics that identify edited regions inside otherwise real images. Scam AI flags both shallowfakes and fully AI-generated media, and tells the adjuster which kind it found.

Q: Does fraud detection slow down legitimate claims?
A: Media screening runs automatically between upload and adjudication, so genuine claims are not delayed — a clean result keeps the claim on the straight-through path with no extra step for the claimant. The effect is usually the opposite of friction: because fabricated and manipulated media is filtered out with evidence attached, adjusters spend less time on manual suspicion and more legitimate claims qualify for fast-track handling. Fraud detection at intake is what makes higher automation rates safe.

## Meeting deepfake protection
URL: https://scam.ai/solutions/meeting-security

Continuous verification of who is really on the call — powered by Halo.

Key stat: $25M — stolen in a single deepfaked video call — every 'participant' but the victim was synthetic (source: Arup case, 2024)

### The threat

One synthetic participant on one video call can move millions. Deepfaked vendors, cloned colleagues, and replayed backgrounds are already inside enterprise meetings.

### What Halo does during the call

The check runs the whole way through the meeting, on the participant's device, not just at the door.

[Halo](https://scam.ai/halo) verifies every face and background with Eva V1.6, and raises a risk signal the moment something turns synthetic.

### Verified at join is not verified at minute forty

A call is live: a participant can hand off, an account can be compromised, or a synthetic face can switch on right when the meeting turns to money. Halo holds the result for the whole call, not just the join.

### FAQ

Q: How can you tell if someone on a video call is a deepfake?
A: Human tells — odd blinking, blurring at the edge of the face, lag between expression and movement — are increasingly unreliable, because modern face swaps have eliminated most visible glitches. Reliable identification requires algorithmic analysis of the video itself: generation artifacts, face-boundary inconsistencies, and signals invisible to a participant watching a compressed stream mid-conversation. Halo runs that analysis continuously on-device during the call and surfaces a risk signal when a face is synthetic, taking the judgment off the busy participant.

Q: What is on-device deepfake detection and why does it matter for meetings?
A: On-device detection means the analysis runs locally on the participant's machine instead of sending call video to the cloud. For meetings this matters twice: sensitive discussions — deals, legal matters, board business — never leave the device to get a result, and detection works in real time without a network round trip. Halo runs Eva V1.6 locally, verifying faces and backgrounds continuously. Scam AI is SOC 2 Type II audited and GDPR compliant; with Halo, the meeting content is never shipped anywhere.

Q: What was the $25M deepfake video call incident?
A: In the 2024 Arup case, a finance employee joined a video call to discuss a confidential transaction. Every other participant — including the CFO who authorized the payment — was a deepfake. Initially suspicious of an email request, the employee was reassured by familiar faces on video and transferred roughly $25M across multiple payments. It is the canonical example of why video now needs verification: being on the call is no longer proof of who is on the call.

## Pig butchering scams
URL: https://scam.ai/solutions/pig-butchering

Detect the fake faces and forged proof behind long-con romance and investment scams — before your users are drained.

Key stat: $5.8B — reported U.S. losses to crypto-investment ('pig butchering') scams in 2024 (source: FBI IC3)

### The threat

Scammers spend weeks building trust with a fabricated persona, using AI-generated profile photos and deepfake video calls, then steer the victim into a fake investment platform. By the time the money moves, the 'person' never existed.

### Detection at every stage of the con

Pig butchering touches several products before money moves, and Scam AI scores the media at each one through the REST API: profile photos at signup, video-call frames for face swaps, and the fake dashboards, statements, and payment proofs shown to victims. Each result is a separate chance to break the con before the savings leave.

### Weeks of runway means weeks of chances to intervene

Pig butchering is slow by design, and that patience is its weakness: every stage leaves synthetic media behind, from the generated profile photo to the deepfaked call to the fake trading dashboard and the forged withdrawal-fee invoice. Detecting any one of them interrupts the con while the victim has lost little or nothing (the FBI counted $5.8B in reported U.S. losses in 2024).

### FAQ

Q: What is a pig butchering scam?
A: Pig butchering is a long-con investment scam: the fraudster builds a weeks-long relationship with the victim — often romantic, started on a dating app or through a wrong-number message — then introduces a fake investment platform showing fabricated returns. The victim invests more, fattening the pig, until the scammer disappears with everything. The persona is synthetic, built on AI-generated photos and deepfaked video calls. The FBI attributed $5.8B in reported U.S. losses to these crypto-investment scams in 2024.

Q: How can platforms detect romance scammers using AI-generated photos?
A: AI-generated profile photos carry statistical artifacts that detection models identify even in flawless-looking images — inconsistencies in how generators render texture, lighting, and facial detail. Screening at signup catches synthetic personas before they contact anyone, and rescreening when an account shows scam-pattern behavior catches personas that changed photos later. Scam AI's REST API returns a scored result per image, so trust and safety teams can gate, review, or ban on thresholds they control instead of waiting for user reports.

Q: How can banks stop pig butchering before the money is gone?
A: Banks usually meet pig butchering at the transfer, when a customer moves unusual sums to crypto exchanges or unfamiliar accounts. The intervention problem is that victims defend the scam — they believe the investment is real. Forensic evidence changes those conversations: screening the proof the victim was shown — platform screenshots, account statements, payment confirmations — produces a manipulation result a banker can show the customer. Proving the document is forged is more persuasive than saying the relationship is fake.

## Deepfake social engineering attacks
URL: https://scam.ai/solutions/social-engineering

Catch the deepfaked faces attackers use to manipulate employees into transfers, access, and approvals.

Key stat: 3,000% — growth in deepfake-based fraud attempts in a single year (source: Onfido Identity Fraud Report)

### The threat

Modern social engineering wears a real face: a deepfaked executive on a video call, or an urgent message that looks like it came from someone you trust, pressuring an employee to move money or hand over access. In the 2024 Arup case, one such call moved $25M.

### A verification layer across your riskiest requests

Social engineering converges on a few moments like a payment approval, a credential reset, or an access grant, so Scam AI deploys there: employees who handle them run Halo to verify faces in every meeting on-device. Any media used to build trust is scored through the REST API, and a flagged face or forged document pauses the request for verification through a known channel.

### Awareness training assumed the fake would look fake

A decade of training taught employees to hunt for tells like bad grammar and video that looks off, but generative AI retired those tells (Onfido measured 3,000% single-year growth in deepfake fraud attempts). Training still matters for the instinct to slow down and verify, but the human eye can no longer be the detection layer: people own pausing and verifying out-of-band, software owns judging whether the face is real.

### FAQ

Q: What is deepfake social engineering?
A: Deepfake social engineering is manipulation that borrows a trusted face: an attacker appears on video as an executive, colleague, or vendor and uses that trust to push an employee into a transfer, a credential reset, or an access grant. It extends classic pretexting with synthetic media convincing enough to defeat visual judgment, stacking urgency and confidentiality to keep the target from verifying. Defenses combine process — out-of-band verification for sensitive requests — with in-meeting detection that flags the fake directly.

Q: Can employees be trained to spot deepfakes?
A: Training helps with process but not with perception. Employees can reliably learn to pause on urgent requests, verify through separate channels, and escalate anomalies — behaviors that stop many attacks however good the fake. What they cannot learn is to see deepfakes: modern face swaps eliminate the glitches older awareness material taught people to find. Effective programs pair the trained behaviors with algorithmic detection, such as Halo running during meetings, so identifying the synthetic face is not a human responsibility.

Q: How should a company verify an urgent request made over video?
A: Urgent video requests should be verified through a channel the requester did not control: a phone number from the directory, a message in the corporate chat thread, or an in-person check — never contact details provided during the call. The policy works best unconditional, applied to every payment or access request above a threshold, because attackers engineer exceptions to feel justified. In-meeting detection strengthens the policy by flagging synthetic faces during the call, giving employees grounds to invoke it.

## HR / Hiring
URL: https://scam.ai/solutions/hiring-fraud

Protect hiring end to end — make sure the person you interview is the person you hire, not a deepfake stand-in or a proxy.

Key stat: 1 in 4 — candidate profiles worldwide projected to be fake by 2028 (source: Gartner)

### The threat

Remote hiring has a new fraud: a real interviewee replaced by a deepfake, or a proxy standing in for the candidate who was screened. It is a growing vector for insider risk and sanctioned-actor infiltration.

### From first interview to first day

Detection covers the two points where identity can be swapped: Halo flags a swapped or lip-synced face live in the interview, and the REST API confirms at onboarding that the hire matches who interviewed. Flagged interviews follow a set path of re-interview, escalation, or withdrawal, while genuine candidates notice nothing.

### How the fake candidate actually gets through

The usual pattern is a real person interviewing behind a real-time face swap, or a proxy answering off-screen while someone else shows up to work. The stakes reach sanctioned-actor infiltration with code and payroll access, and the interview is the last moment the fraud must perform live (Gartner projects one in four candidates will be fake by 2028).

### FAQ

Q: How do fake candidates use deepfakes in remote job interviews?
A: Fake candidates run real-time face-swap software fed into the call through a virtual camera, so the interviewer sees a fabricated face that moves naturally — driven by a real person who is not who they claim to be. Variants include lip-sync overlays and proxy setups where an off-screen expert supplies answers. The goal is to pass assessment as one identity and place a different person, or a sanctioned actor, into the role. Detection during the interview flags the swap mid-conversation.

Q: How can recruiters verify a remote candidate is real?
A: Recruiter-level checks — asking the candidate to turn their head, wave a hand across their face, or adjust the lighting — can disrupt some older face swaps, but current tools handle these prompts well, and the checks make interviews adversarial for genuine candidates. Reliable verification pairs the interview with on-device detection: Halo analyzes the candidate's video continuously and flags swapped or rendered faces without the choreography. Identity verification at onboarding then confirms the new hire matches the person who interviewed.

Q: Why is hiring fraud a security problem and not just an HR problem?
A: A fraudulent hire is an insider threat with a badge: day one brings credentials, source code access, customer data, and a paycheck that can flow to a sanctioned entity. Documented cases include remote IT workers placed by hostile states inside Western companies this way. So the response spans recruiting and security together — screening interviews for deepfakes, verifying identity at onboarding, and treating a flagged candidate as a security event with an escalation path, not merely a rejected application.

## Content moderation
URL: https://scam.ai/solutions/content-moderation

Score uploads for AI generation and manipulation before synthetic content spreads.

Key stat: 8M — deepfakes projected to be shared online in 2025 — up from 500k in 2023 (source: World Economic Forum)

### The threat

Synthetic media arrives faster than human moderation can review it: deepfaked public figures, AI-generated abuse material, and coordinated fake content at platform scale.

### One API call between upload and publish

Detection slots into the ingest pipeline you already run: each uploaded image or video is scored by the REST API before it goes live, with the evidence behind it. Auto-action at high scores and queue humans at medium, on thresholds your policy team tunes per surface.

### Detection powers labeling, not just takedowns

Policy is no longer allow-or-remove, since much AI content is legitimate and the real duty is disclosure. A scored result lets platforms label routine synthetic content, restrict synthetic media of real people, and remove deceptive deepfakes (the World Economic Forum projected 8M deepfakes shared online in 2025).

### FAQ

Q: How do platforms detect AI-generated content at scale?
A: At platform scale, detection runs as an automated scoring step in the upload pipeline: every image and video is analyzed for generation artifacts, face swaps, and manipulation before publication, and the score feeds moderation rules as spam and abuse signals already do. Human review is reserved for the contested middle. Scam AI's REST API is built for this pattern, with real-time results at upload, batch endpoints for library rescans, and evidence attached to each result for moderator review and transparency reporting.

Q: Can deepfake detection keep up with upload volume in real time?
A: Real-time detection at upload volume is an architecture question: the check must return a result within the latency budget of the publish flow and scale with traffic. Scam AI's API is designed for both modes — synchronous scoring for the upload path and batch processing for backfills and rescans — so platforms do not choose between coverage and speed. The practical pattern is scoring everything at ingest, auto-actioning only at high scores, and letting the moderation queue absorb the borderline cases.

Q: Should AI-generated content be removed or labeled?
A: Most platforms are converging on tiered enforcement rather than blanket removal: benign synthetic content stays up with a label, synthetic depictions of real people face restrictions, and deceptive deepfakes — fabricated statements, non-consensual imagery, fraud content — are removed. Every tier depends on knowing the content is synthetic, which self-disclosure alone cannot provide. Automated detection supplies that knowledge at scale, and results with evidence let platforms defend both the label and the takedown in appeals and transparency reports.

## Misinformation detection
URL: https://scam.ai/solutions/misinformation

Detect the deepfakes and synthetic media behind misinformation before it spreads and erodes trust.

Key stat: 90% — of online content could be AI-generated by 2026 (source: Europol Innovation Lab)

### The threat

Deepfaked officials, fabricated events, and AI-generated media spread faster than they can be debunked. The 'liar's dividend' also lets bad actors dismiss real footage as fake.

### Results at the speed of the news cycle

Newsrooms and fact-checkers submit suspect media to the REST API and get back a scored, evidence-backed result fast enough for the piece being written now. Because each result names the signals that fired, the output is citable.

### Verifying the real is half the job

Once the public knows video can be faked, anyone caught on genuine footage can call it synthetic, so detection has to run both ways. A clean result is as valuable as a flag, letting a newsroom stand behind authentic footage (Europol projects up to 90% of online content could be AI-generated by 2026).

### FAQ

Q: How do fact-checkers verify whether a video is a deepfake?
A: Fact-checkers combine traditional verification — sourcing the original upload, checking location and timing, contacting the people depicted — with forensic analysis of the media, examining faces and frames for generation artifacts and manipulation traces that survive compression. The forensic layer matters most when a fake is well made and context checks are inconclusive. Scam AI provides that layer through the REST API: a scored result with the signals that fired, delivered fast enough to inform coverage before the content peaks.

Q: What is the liar's dividend?
A: The liar's dividend is the benefit bad actors gain from the existence of deepfakes: because the public knows video can be fabricated, genuine footage can be dismissed as fake. The term describes a second-order harm of synthetic media — beyond deceiving people with fakes, it corrodes the evidentiary value of everything real. Countering it requires the ability to affirm authenticity, not just flag manipulation — so results that clear genuine media matter as much as results that catch fakes.

Q: Can detection tools confirm that footage is authentic?
A: Detection models score media in both directions: a result can flag signs of AI generation or manipulation, or report none are present, with evidence behind either conclusion. A clean forensic result is not absolute proof, but combined with provenance and sourcing it gives newsrooms and platforms grounds to treat footage as genuine — and resist pressure to retract real material dismissed as fake. Scam AI attaches the fired signals to every result so the reasoning is citable, not just the score.

## Deepfake brand protection
URL: https://scam.ai/solutions/brand-protection

Catch the deepfakes, fake ads, and impersonations that hijack your brand and your executives' likenesses.

Key stat: $2.9B — reported consumer losses to impersonation scams in 2024 (source: FTC)

### The threat

Attackers clone your executives, run fake ads carrying your logo, and spin up impersonation accounts, driving fraud that looks like it came from you.

### From suspicious sighting to takedown-ready evidence

Send suspect media to the REST API and get back a result on whether an executive's face is generated or an ad shows AI manipulation, with the signals behind the call. Confirmed fakes move into your takedown workflow with forensic evidence attached.

### Every fake ad spends your trust to scam your customers

Brand impersonation is a double theft: your customers lose money to a fake fronted by your CEO's face, and your brand absorbs the durable damage as every victim ties the scam to your name. Speed is the defense that matters, since a fake that runs for weeks instead of hours is measured in victims, and detection with evidence compresses the time from sighting to takedown (the FTC counted $2.9B in reported impersonation losses in 2024).

### FAQ

Q: How do companies detect deepfake ads using their executives?
A: Deepfake ad campaigns are detected through a combination of monitoring and forensics: social listening and ad-library monitoring surface suspect creative, and detection models then confirm whether the executive's face is genuine or AI-generated. The forensic step matters because takedown requests supported by evidence move faster than bare assertions. Scam AI scores suspect media through the REST API, returning results with the signals that fired so brand and legal teams can act immediately.

Q: How do you get a deepfake impersonation taken down?
A: Takedowns run through the platform hosting the content — social networks, ad networks, app stores, or registrars — and speed depends on the quality of the report. A request with forensic evidence the media is AI-generated, alongside proof of the impersonated party's identity, is harder to deprioritize than a bare complaint. Scam AI results attach that evidence: the manipulation score and the signals that fired. Companies route confirmed fakes into a prepared takedown workflow so the response takes hours, not weeks.

## AI phishing & spam detection
URL: https://scam.ai/solutions/phishing-spam

Screen the images, documents, and media inside phishing and spam campaigns for AI generation and manipulation.

### The threat

Phishing and spam increasingly carry AI-generated images, fake invoices, and synthetic identities to slip past filters and trick recipients into paying, clicking, or handing over credentials.

### A media forensics layer behind your mail filter

Scam AI gives your mail gateway a signal it cannot produce: attachments and embedded media are scored by the REST API for AI generation and tampering. That score joins your filter's features, so a clean-looking message carrying a forged invoice still gets caught.

### The lure moved from text into images

Filters read text but mostly skip images, so the lure moved into media: a pixel-perfect PDF invoice, a payment instruction hidden in a screenshot. Scoring the media itself catches the forged document by the forensic traces of its manufacture.

### FAQ

Q: How is AI-generated phishing different from traditional phishing?
A: Traditional phishing was detectable by its flaws: template reuse, crude forgeries, and language errors that filters and users learned to catch. AI-generated phishing removes the flaws and adds variation — every message and forged document can be unique, defeating signature and hash-based blocking, and lures arrive as convincing images and PDFs rather than suspicious text. Detection shifts accordingly: instead of matching known-bad patterns, filters need forensic signals on whether the media was generated or manipulated — the layer Scam AI adds.

Q: Can fake invoice attachments be detected automatically?
A: Forged invoices carry detectable traces even when they look perfect: editing artifacts, font and layout inconsistencies, and signs of AI generation. Automated document forensics finds these at scale — every attachment scored as it arrives, not left to an accounts-payable clerk to notice. Scam AI's REST API returns a per-document result with evidence, which mail security stacks consume as another signal: a message with a synthetic invoice gets quarantined or flagged even when its text and sender pass every other check.

Q: What is image-based phishing and why do filters miss it?
A: Image-based phishing puts the deception inside media instead of text: a screenshot styled as a security alert, a QR code replacing a clickable link, a payment instruction rendered as a picture. Filters tuned to message text and URLs miss it because such messages carry minimal, innocuous text — the attack lives in pixels the classifier never reads. Countering it requires analyzing the media for manipulation and synthetic origin — a media forensics signal alongside conventional filtering closes the gap.

# Industries

## KYC/AML compliance
URL: https://scam.ai/industries/kyc-aml-compliance

Meet KYC and AML obligations against AI-era fraud — deepfake selfies, forged IDs, and synthetic identities caught at onboarding and beyond.

Key stat: $4.6B — in AML and KYC enforcement actions issued globally in 2024 (source: Fenergo)

### Compliance under AI-era fraud

Regulators expect you to keep synthetic identities and manipulated documents out of onboarding. Scam AI scores the selfie for deepfakes, verifies liveness, and checks every submitted document for forgery — so your KYC/AML program holds up against generative-AI attacks.

### Auditable by design

Every result ships with the signals that fired, giving compliance teams and auditors defensible evidence for approve, decline, and enhanced-due-diligence decisions — not an opaque score.

### What examiners expect now that IDs can be generated

KYC rules assumed the hard part was matching a person to a document; generative AI broke that assumption. FATF guidance asks whether your identity-proofing resists spoofing, the Bank Secrecy Act's Customer Identification Program rule requires a reasonable belief in each customer's true identity, and FinCEN's 2024 alert lists the generative-AI red flags institutions should detect and report. Scam AI answers all three with documented, testable detection of deepfake selfies and forged documents, with evidence trails built for examiner review.

### Where detection sits in the verification stack

Eva V1.6 slots into your existing stack as a dedicated synthetic-media layer, not another all-in-one IDV tool: it scores the selfie for AI generation before the biometric match runs — a match against a fake face is still a fake — and runs forensics on every ID and supporting document. The same checks re-run at enhanced due diligence and periodic refresh, so a customer who onboarded legitimately cannot later swap in synthetic media unnoticed. With $4.6B in AML and KYC enforcement actions issued globally in 2024, an audit trail that carries reasons rather than bare scores is the point.

### FAQ

Q: How are deepfakes used to bypass KYC checks?
A: Fraudsters attack KYC at two points. At the selfie step, they use face swaps, AI-generated faces, or replayed synthetic video to pass liveness and biometric matching — often injecting the fake feed into the session with a virtual camera. At the document step, they submit AI-generated or template-forged IDs and proof-of-address documents. Because both artifacts agree with each other, the fake identity sails through checks that only match a face to a document; detection has to interrogate the media itself.

Q: Do regulators require deepfake detection for KYC and AML programs?
A: No rule names a specific technology, but the expectations point one way. The U.S. Customer Identification Program rule requires a reasonable belief in a customer's true identity, FinCEN's 2024 alert asks institutions to identify and report deepfake-related red flags, and FATF guidance tells firms to evaluate whether their identity-proofing resists spoofing. If synthetic identities are passing your onboarding, examiners will ask what controls you had — documented synthetic-media detection is the clearest way to show your program kept pace.

Q: What evidence does a compliance team get with each detection result?
A: Every Scam AI result includes the signals that drove it: which regions of a face or document showed manipulation, what class of generation or tampering was detected, and the score behind the verdict. An analyst reviewing a declined onboarding can see why, an MLRO can cite indicators in a suspicious activity report, and an auditor can trace any decision to evidence. Results are retained while the underlying media is discarded after scoring — a complete audit trail without a data-retention liability.

## On-device deepfake detection agent
URL: https://scam.ai/industries/on-device-agent

Real-time detection that runs on the device — media and identity verified locally, with nothing leaving the endpoint. Shipping today in Halo; an embeddable agent for your own apps and devices is in development.

Key stat: 1 / 5 min — a deepfake fraud attempt occurred every five minutes in 2024 (source: Entrust Identity Fraud Report)

### Detection at the edge

The same Eva V1.6 detection runs locally on the endpoint: deepfakes, forged documents, and manipulated media flagged in real time without a round trip to the cloud. Halo ships this today for live video meetings. Packaging the same engine as an embeddable agent your own app or device can call is in development, alongside the native SDKs on the API roadmap.

To run it today, [see Halo](https://scam.ai/halo).

### Private and always on

Because analysis happens on the endpoint, sensitive media never leaves the device — ideal for regulated environments, offline scenarios, and low-latency flows where cloud calls aren't an option.

### Data minimization as a compliance position

To check media for manipulation, most architectures ship it to someone's cloud — and for regulated deployments that round trip is itself a compliance event, raising GDPR questions about data minimization, purpose limitation, and cross-border transfer. On-device detection removes the event entirely: Eva V1.6 runs locally and only the result leaves the device, so a privacy impact assessment has no new data flow to assess. Scam AI is SOC 2 Type II audited and GDPR compliant, and your users' faces never become someone else's dataset.

### Where an edge detection agent fits

The embeddable agent is being built for capture-time checks inside banking and fintech apps, kiosks and retail counters that need to verify the person standing there, and field or regulated endpoints where data cannot leave the machine. Today the shipping path for on-device detection is Halo on the participant's own machine, with the REST API covering every other integration in the meantime. If one of these patterns is yours, tell us — early access shapes what we build first.

### FAQ

Q: What is on-device deepfake detection?
A: On-device deepfake detection runs the model on the endpoint — a phone, kiosk, laptop, or embedded device — instead of sending media to a cloud API. The device captures a face image, document photo, or video frame; the local model scores it for AI generation and manipulation; the app gets a result in real time. Because analysis happens where media is captured, nothing sensitive leaves the device: lower latency, offline operation, no new data flow for privacy review.

Q: Does the on-device agent work without an internet connection?
A: Yes — that is the point of running inference on the endpoint: results do not depend on connectivity. Halo already works this way, analyzing meeting video locally on the participant's machine. The same property is what makes the embeddable agent, currently in development, suited to kiosks in low-coverage locations, field workflows like claims adjusting or benefits processing in remote areas, and any environment where the network is untrusted or intermittent. Model updates arrive when the device connects, and result metadata can queue for later sync if you want centralized reporting.

Q: How is an on-device agent different from calling a detection API?
A: Both return the same thing: a manipulation result with signals from the Eva V1.6 model. The difference is where the media goes. With the REST API — available today, in any language — your backend sends media to Scam AI, which analyzes, scores, and discards it; right for server-side pipelines like upload moderation or claims processing. On-device, the model runs on the endpoint instead, so media never leaves it: today that ships as Halo for live meetings, and as an embeddable agent for your own apps once it is out of development. Teams choose on-device for privacy-sensitive capture, offline operation, and latency budgets.

## Deepfake detection for financial services
URL: https://scam.ai/industries/financial-services

Stop synthetic identities at onboarding and deepfaked executives on video calls before the wire is released.

Key stat: $40B — projected U.S. fraud losses enabled by generative AI by 2027 (source: Deloitte)

### Where banks get hit

KYC bypass with rendered faces, forged statements in lending, and deepfaked executives authorizing transfers.

### How Scam AI helps

Eva V1.6 screens enrollment selfies and document uploads for AI generation and tampering, and Halo flags deepfaked participants in the meetings where transfers get approved — every result with evidence attached.

### One platform across the bank

From retail onboarding to lending-document intake to executive wire approvals, the same detection covers every trust decision — no separate point solution to buy and wire up per channel.

### Mapping detection to FinCEN and BSA expectations

FinCEN's late-2024 alert warned institutions about deepfake media used to defeat identity verification, and reminded them that suspected generative-AI use belongs in suspicious activity reporting — putting synthetic-media detection inside existing Bank Secrecy Act obligations. Eva V1.6 scores enrollment selfies and uploaded documents for AI generation and manipulation, and every result ships with the signals that fired: concrete language for SAR narratives instead of a hunch. Scam AI is SOC 2 Type II audited and GDPR compliant, so the detection layer does not create a new audit finding.

### From account opening to wire release

Detection only reduces losses where money actually moves: account opening, where Eva V1.6 screens the applicant's selfie and ID before provisioning; the lending pipeline, where pay stubs and statements get forensics before underwriting sees them; and high-value approvals, where Halo flags synthetic video in the meeting authorizing the transfer. Each check is a REST API call that drops into your existing orchestration, so fraud teams keep their case-management tools and decision engines. Deloitte projects $40B in U.S. fraud losses enabled by generative AI by 2027 — the institutions instrumenting these checkpoints early set the loss curve.

### FAQ

Q: How do banks detect deepfakes during digital account opening?
A: Banks layer synthetic-media detection on top of standard identity verification. When an applicant submits a selfie and ID, a detection model like Eva V1.6 analyzes the face for AI generation, face swapping, and presentation attacks, while document forensics checks the ID and supporting paperwork for tampering and generated content. The result is a verdict with a score and evidence that feeds the bank's decision engine: genuine applicants pass in seconds, and suspicious enrollments route to manual review before an account is opened.

Q: What did FinCEN say about deepfakes, and what should banks do about it?
A: FinCEN's 2024 alert warned that criminals are using deepfake media, including AI-generated identity documents and altered photos, to open accounts and defeat verification. It asked institutions to watch for red flags like inconsistencies between a customer's photo and their document, and to reference generative-AI indicators in suspicious activity reports. Practically, banks should add synthetic-media scoring at onboarding and document intake, with detection tooling that produces explainable evidence for SAR narratives and examiners.

Q: Can deepfake detection integrate with our existing KYC vendor?
A: Yes. Scam AI sits alongside, not in place of, your identity verification stack. The REST API accepts the same selfie and document images your KYC vendor already captures and returns a manipulation result with the signals that fired. Most teams call it from their orchestration layer as an additional onboarding check, using the score to trigger step-up verification or review. No rip-and-replace: your IDV provider keeps matching identities, and Scam AI answers the separate question of whether the media itself is real.

## Deepfake detection for the public sector
URL: https://scam.ai/industries/public-sector

Protect officials and citizens from impersonation campaigns.

Key stat: $2.9B — lost to impersonation scams — including fake government agencies — in 2024 (source: FTC)

### Where government gets hit

Deepfaked officials spreading disinformation, forged documents in benefits claims, and synthetic identities defrauding public programs at scale.

### How Scam AI helps

Eva V1.6 scores documents at benefits intake and verifies media for press teams; Halo protects staff meetings on-device. Every result ships with the evidence behind it.

### Trust in public communication

When an official's likeness can be faked in minutes, verified media keeps citizens able to trust what their government actually said — and to spot what it didn't.

### Impersonation now has its own rulebook

Impersonating a government official has always been a federal crime; generative AI made it scalable, and in 2024 the FTC's Government and Business Impersonation Rule took effect to match. NIST's digital identity guidelines (SP 800-63) contemplate exactly the presentation and injection attacks that deepfake selfies and forged documents represent. Scam AI covers both directions — media verification to protect citizens from fake versions of you, deepfake and document forensics at intake to protect programs — each producing evidence an oversight body can review.

### From benefits intake to the press office

At benefits intake, Eva V1.6 runs document forensics on identity documents and income statements before eligibility decisions are made — the same checkpoint where synthetic identities drained pandemic-era programs. In the press office, media verification gives comms teams a defensible answer when a video of an official surfaces, before the clip finishes its first news cycle. For sensitive coordination, Halo runs on staff machines during video meetings without sending meeting content anywhere.

### FAQ

Q: How can a government agency verify whether a video of an official is real?
A: Run the clip through a detection model built for synthetic media. Scam AI's Eva V1.6 analyzes the video frame by frame for face swapping, AI generation, and manipulation, and returns a result with the signals that drove it. That gives a communications team something a press statement can stand on: not just "we believe this is fake," but documented detection evidence, available in minutes.

Q: How does deepfake detection reduce benefits and public-program fraud?
A: Most program fraud enters through intake: forged identity documents, AI-generated pay stubs and statements, and synthetic identities assembled to pass remote verification. Detection at intake scores every document and selfie for generation and tampering before an eligibility decision is made, routing suspicious cases to investigators with evidence attached. That shifts fraud control from pay-and-chase to prevention at the front door, and aligns with the identity-assurance expectations in NIST's digital identity guidelines that most U.S. agencies build their remote proofing around.

Q: Can public-sector deployments keep media on their own infrastructure?
A: Yes. For meeting protection, Halo runs entirely on the endpoint: detection happens on the staff member's machine and meeting content never leaves it. For intake and verification pipelines, the on-device agent brings the same Eva V1.6 model inside your boundary, so document and face analysis runs locally where policy prohibits external services. Where the REST API is used, media is analyzed, scored, and discarded (Scam AI retains results, not content), and the company is SOC 2 Type II audited and GDPR compliant.

## E-commerce & marketplaces
URL: https://scam.ai/industries/e-commerce-marketplaces

Keep fake sellers and cloned storefronts out.

Key stat: $48B — global e-commerce losses to online payment fraud in 2025 (source: Juniper Research)

### Where marketplaces get hit

Synthetic seller identities, AI-generated product imagery for goods that don't exist, forged business documents, and cloned storefronts harvesting buyers.

### How Scam AI helps

Seller onboarding liveness, document verification on business credentials, and image forensics on listings — trust signals for every side of the marketplace.

### Trust for both sides

Buyers meet real sellers and genuine listings; sellers reach a marketplace that isn't overrun by clones and fakes — protection for the whole transaction, not just one side of it.

### Seller verification is law now, not best practice

In the U.S., the INFORM Consumers Act requires marketplaces to collect and verify bank, tax, and contact details from high-volume third-party sellers; in the EU, the Digital Services Act's traceability-of-traders provisions require verifying trader information before they can sell. Both regimes assume the documents sellers submit are checkable — exactly what generative AI undermines, now that a fraud ring can fabricate a matching set of registration, ID, and bank documents. Eva V1.6 runs forensics on seller credentials and deepfake screening on verification selfies, so what you are legally required to verify is tested for synthesis, not just collected.

### Detection across the marketplace lifecycle

Marketplace fraud is not a single event, so detection is not a single checkpoint: deepfake and document screening at seller onboarding, image forensics on listing photos to catch goods that do not exist, and re-verification at payout redirection or credential recovery to close the account-takeover path. In disputes, media checks on buyer-submitted damage and return photos separate genuine claims from refund abuse. Each is one REST API call — and with Juniper Research putting 2025 e-commerce losses to online payment fraud at $48B, prevention compounds at these moments.

### FAQ

Q: How do online marketplaces detect fake sellers?
A: Layered verification at onboarding. The applicant's selfie is screened for deepfakes and AI-generated faces, their identity and business documents get forensic analysis for forgery and generation, and the results feed the marketplace's risk decision before selling privileges are granted. It matters because fraud rings submit internally consistent document sets that pass visual review. Detection models like Eva V1.6 examine the media itself (generation artifacts, tampering traces, template reuse), so a fabricated seller package gets caught even when the documents agree.

Q: Can AI-generated product images be detected on listings?
A: Yes. Image forensics can identify product photos AI-generated rather than captured, plus composites and manipulations. That signal matters for two abuse patterns: listings for goods that do not exist, where the storefront is synthetic and buyers pay for nothing, and counterfeit listings dressed up with generated imagery to dodge duplicate-image matching against the original store. Marketplaces score listing images through the REST API at publish time, gating high-risk listings for review before they go live, not after buyer complaints arrive.

Q: What does the INFORM Consumers Act require marketplaces to verify?
A: The INFORM Consumers Act requires marketplaces to collect and verify information from high-volume third-party sellers (bank account details, government-issued ID or tax documentation, and contact information) and to suspend noncompliant sellers. It also mandates disclosure of seller information to buyers past thresholds. The catch: collection is easy, and verification is where fraud lives; AI-generated IDs and fabricated business documents satisfy a checkbox process. Document forensics and deepfake screening make the verification step substantive, which is what the statute asks for.

## Dating apps
URL: https://scam.ai/industries/dating-apps

Keep fake profiles, AI-generated photos, and deepfaked video calls off your platform — before they turn into romance scams.

Key stat: $1.1B — reported U.S. losses to romance scams in a single year (source: FTC)

### Where dating apps get hit

Scammers build trust with fabricated personas — AI-generated profile pictures, deepfaked video calls — then steer matches toward money or fake investments. The whole con depends on the fake looking real.

### How Scam AI helps

Screen profile photos for AI generation at signup, check verification selfies for deepfakes and presentation attacks, and surface synthetic-persona risk signals early — protecting users while the con is still warming up.

### Safety duties are shifting from optional to statutory

The UK's Online Safety Act puts duties on user-to-user services to assess and mitigate illegal content, and romance fraud — a crime built on fake personas — sits squarely inside that duty, while the EU's Digital Services Act requires platforms to act on illegal content and be transparent about moderation. Every one of these regimes presumes the platform can tell fabricated personas from real ones. Scam AI supplies that: AI-generation screening on profile photos, deepfake and presentation-attack detection on verification selfies, and risk signals your trust and safety team can document.

### From signup to verified profile

At signup, every profile photo is scored for AI generation — one REST API call in your upload path — catching the fabricated faces that anchor scam personas before a profile reaches the match queue. At verification, Eva V1.6 screens the badge selfie for deepfakes and presentation attacks, so the badge means what users think it means instead of laundering a fake into credibility. Real-time screening inside your own in-app video calls needs the embeddable on-device agent, still in development; today the checkpoints are photo upload and verification.

### FAQ

Q: How do dating apps detect fake profiles that use AI-generated photos?
A: By scoring photos at upload rather than waiting for user reports. Generated faces carry statistical artifacts that detection models identify even when the image looks flawless. Scam AI's Eva V1.6 screens every profile photo at signup and on photo changes, returning a result before the profile enters matching. Front-door screening matters because AI-generated faces defeat the older defense: reverse-image search only catches photos stolen from somewhere else. A generated face has no source to find; detection must examine the image itself.

Q: Can deepfakes be detected on live video calls inside a dating app?
A: Real-time in-call screening is what the loophole demands — victims are told to "insist on a video call," and scammers answer with face-swapped video that passes inspection — but embedding it inside your own video stack requires the on-device agent, which is in development. What runs today is screening at the checkpoints you already control: profile photos at upload and verification selfies, both scored through the REST API for face swaps and presentation attacks. Platforms use those signals to gate a persona for review, warn the other user, or feed the account's risk score.

Q: What does the UK Online Safety Act mean for dating platforms?
A: The Online Safety Act puts duties on user-to-user services, dating apps included, to assess the risk of illegal content and take proportionate measures against it, with romance fraud among the harms in scope. Larger categorized services face expectations to offer identity-verification options. Practically, a platform must demonstrate systems that find fabricated personas: photo screening for AI generation, verification selfies checked for deepfakes, and documented enforcement. Detection results with evidence support the mitigation itself and the risk-assessment paperwork the regime requires.

## Telecom fraud detection
URL: https://scam.ai/industries/telecom

Protect SIM changes, account access, and support channels from synthetic identities and social engineering.

Key stat: $39B — estimated annual global losses to telecom fraud (source: CFCA Global Fraud Loss Survey)

### Where telecom gets hit

A SIM swap starts with a story: an attacker armed with breached personal data talks a support rep or retail associate into moving a victim's number, then intercepts the one-time codes protecting their bank, email, and crypto accounts. Port-out fraud, account-recovery takeovers, and synthetic identities at activation follow the same pattern — the human check is the weakest link.

### How Scam AI helps

Scam AI re-verifies the customer is live and real on high-risk changes — a face-liveness check the fraudster on the line can't pass — and surfaces device and document risk signals before a SIM or account moves.

### The FCC has rewritten the SIM-swap playbook

The FCC's 2023 rules on SIM swapping and port-out fraud require wireless providers to authenticate customers securely before processing a SIM change or number port, and to notify the customer immediately — flexible on method, accountable for outcomes. They sit on top of long-standing CPNI obligations penalizing weak account-access controls, leaving one open question: what "secure authentication" means once generative AI defeats knowledge-based checks and naive photo review. Scam AI's answer interrogates the media itself — face liveness and deepfake screening on high-risk changes, document forensics on forged IDs in recovery flows.

### Where detection sits in the SIM-change flow

The same entry points recur across the CFCA's estimated $39B in annual telecom fraud losses: SIM swaps, number ports, account recovery, and new-line activation. When a SIM change or port-out arrives — in-app, in store, or through support — the flow triggers a step-up face check screened by Eva V1.6 for deepfakes and presentation attacks, while identity documents in recovery flows get forensic scoring, so a forged ID fails however polished the story. The REST API drops into order-management and support flows, and the customer notification the FCC requires stays — detection makes the approval behind it trustworthy.

### FAQ

Q: How can carriers prevent SIM-swap fraud?
A: By making the authentication step before a SIM change resistant to the attacks in use. Knowledge-based questions fail because answers leak in breaches; SMS codes fail because the attacker captures the number. The durable control verifies the person: a face check screened for deepfakes and presentation attacks, in the carrier's app or on a retail tablet, plus document forensics on IDs in recovery flows. With the customer notifications the FCC now requires, that closes the social-engineering path into someone's number.

Q: What do the FCC's SIM-swap rules require of wireless providers?
A: The FCC's 2023 rules require wireless carriers to use secure methods to authenticate a customer before executing a SIM change or port-out, to notify the customer immediately on any such request, and to maintain processes for failed authentication and complaints. The rules deliberately avoid prescribing one technology; carriers choose methods that resist current attacks. That flexibility is why providers add face-based verification with deepfake screening on high-risk changes: its strength does not depend on secrets the attacker already holds.

Q: Can identity verification with deepfake detection run in retail stores and the carrier app?
A: Yes, and covering both matters, because fraudsters route to the weakest path. In the app, the SIM-change or recovery flow triggers a face capture that Eva V1.6 screens for deepfakes and presentation attacks before the request proceeds. In stores, the same verification runs on the retail tablet, so a walk-in with a forged ID faces the same check as a remote attacker. The on-device agent runs detection locally on the tablet or phone, keeping biometric media on the endpoint.

# Resources

## Education
URL: https://scam.ai/resources/education

How AI-driven scams work, and how to check for them: tools you can try on your own media, published research, and the rules that apply.

### Check a file yourself

Upload an image and Check if AI tells you whether it looks AI-generated. It is the quickest way to see what detection looks for.

[Try Check if AI](https://scam.ai/platform/check-if-ai)

### Watch yourself become someone else

The Deepfake Playground swaps your face live on your own camera for 30 seconds, so you can judge how convincing a real-time deepfake is on a video call. Nothing is recorded or stored.

[Open the Deepfake Playground](https://scam.ai/platform/deepfake-playground)

### Read the research

Our papers on deepfake detection, document forgery and age-estimation robustness, each summarised by its finding, with the open datasets behind them.

[Browse the research](https://scam.ai/research)

### Know the rules

The regulation tracker follows deepfake and AI-fraud law across the US, Europe and Asia-Pacific (labeling duties, new criminal offenses, election rules).

[Open the regulation tracker](https://scam.ai/resources/regulation)

### Read the team's writing

Posts from the team on how attacks work and how detection handles them.

[Read the blog](https://scam.ai/resources/blog)

### Teach it

Training, certification and curriculum for partner teams, universities and researchers, with lab access to the detection API.

[See the education program](https://scam.ai/resources/education-program)

## Trust Center
URL: https://scam.ai/resources/trust-center

Security, privacy, and compliance documentation.

### Private by design

Media is analyzed, scored, and discarded — we keep the results, not your content. Halo goes further: detection runs on-device and nothing leaves your machine.

### Compliance

Scam AI is SOC 2 Type II audited and GDPR compliant, and a member of C2PA (the Coalition for Content Provenance and Authenticity). We share audit reports and security documentation under NDA; request them below.

### Documents

The policies and terms a security or procurement review asks for, each on its own page.

- [Privacy Policy](https://scam.ai/legal/privacy): What personal data we collect, how we use and protect it, and your rights.
- [Data Processing Addendum](https://scam.ai/legal/dpa): Processor terms for customer personal data under GDPR Article 28.
- [Subprocessors](https://scam.ai/legal/subprocessors): The third parties that help run the service, and how changes to the list are announced.
- [Master Services Agreement](https://scam.ai/legal/msa): The default commercial terms for paid services.
- [Halo EULA](https://scam.ai/legal/halo-eula): The license terms for the Halo on-device app.
- SOC 2 report: Shared under NDA. Ask for it below.

## Changelog
URL: https://scam.ai/resources/changelog

What's new in Scam AI detection — models, APIs, and product.

### Timeline, newest first

- 2026-09-23 — Command-line tool and SDK 0.1.3: Run detection from a terminal with @scam-ai/cli. The Node and Python SDKs and the MCP server move to 0.1.3 the same day. Source: [@scam-ai/cli on npm](https://www.npmjs.com/package/@scam-ai/cli)
- 2026-09-10 — MCP server for AI agents: Gives an AI agent one tool that checks a file, a video or a URL against the detection API, over the Model Context Protocol. Published as @scam-ai/mcp-server. Source: [@scam-ai/mcp-server on npm](https://www.npmjs.com/package/@scam-ai/mcp-server)
- 2026-09-04 — Official SDKs for Node and Python: @scam-ai/sdk for Node and TypeScript and scamai for Python wrap the detection API, so an integration is an install and a key. Source: [@scam-ai/sdk on npm](https://www.npmjs.com/package/@scam-ai/sdk)
- 2026-08-05 — Halo launches for Windows: Real-time deepfake detection for live Zoom, Microsoft Teams and Google Meet calls, running on the device's Snapdragon NPU. No video or audio is sent to the cloud. Source: [Business Wire](https://www.businesswire.com/news/home/20260805515597/en/Scam.ai-Launches-Halo-Real-Time-On-Device-Deepfake-Detection-for-Video-Calls)
- 2026-06-25 — Halo announced with Qualcomm at Computex 2026: An on-device deepfake detection model for live video calls, shown on Qualcomm's booth in Taipei. Source: [Business Wire](https://www.businesswire.com/news/home/20260623132017/en/Scam.ai-Announces-Qualcomm-Partnership-Launches-Halo-Deepfake-Detection-Model-at-Computex-2026)
- 2026-02-02 — No-code detection and self-serve sign-up: Sign up and check media in the browser, with no API integration needed to get started.
- 2026-01-15 — Detection model update: Higher accuracy and faster results, support for 4K video, and stronger audio deepfake detection.
- 2025-12-10 — AI-generated image and video detection: Detection of images and video made by generative AI tools such as Midjourney and DALL·E.
- 2025-11-05 — Audio deepfake detection: Voice-clone detection for MP3, WAV and AAC audio.
- 2025-08-25 — SOC 2 Type I and GDPR compliance: A SOC 2 Type I report on our security controls, and GDPR compliance.
- 2025-01-05 — First public detection model: Our first deepfake detection model, for images and video.

## Affiliate program
URL: https://scam.ai/resources/affiliate

Earn by referring teams that need deepfake and fraud detection.

### Apply and get your link

Apply with a note on your audience and how you would refer teams that need deepfake or fraud detection. Once accepted, you get a tracking link of your own to share.

### Earn on each referral

Every referral through your link is tracked and reported transparently, and you earn a share of the revenue each one generates.

## Partner program
URL: https://scam.ai/resources/partners

Resell, integrate, or build on the Scam AI detection API.

### Who we partner with

Identity platforms, fraud-ops vendors, meeting software, and system integrators embedding detection into their own products and services.

### Three ways to partner

Resellers bring Scam AI detection to their own customers. Integration partners embed the REST API inside their identity, fraud-ops, or meeting products, so results show up in the workflow their users already run. And builders ship new products on the detection engine itself. Tell us which model fits — we scope it on the first call.

## Deepfake detection education program
URL: https://scam.ai/resources/education-program

Training, certification, and curriculum for partners, students, and institutions who want to teach and deploy deepfake detection.

### Who it's for

Implementation partners enabling their teams, universities and bootcamps building media-forensics curricula, and researchers who want structured, hands-on access to detection tooling.

### What's included

Guided courses on deepfakes, synthetic media, and detection; certification tracks; and lab access to the Scam AI API so learners work with real results, not slides.

# Company

## About Scam AI
URL: https://scam.ai/company/about

Scam AI exists because seeing is no longer believing. We build the tools that let people and organizations trust what they see.

### The company

Scam AI — built by Reality Inc. — is a team of researchers and engineers focused on one problem: detecting AI-generated deception at the moment it matters. That work ships as Eva V1.6, our unified detection API for images, video, and audio; Halo, on-device detection for live video calls; and document and ID forgery detection, a separate plan. The company is SOC 2 Type II audited and GDPR compliant.

### Backed by

Scam AI was selected for Berkeley SkyDeck, UC Berkeley's deep-tech accelerator, in 2025, and closed a $2.6M seed round led by Llama Ventures, with participation from SCVxSBI, Welight Capital, SkyDeck Fund, and Beta Fund. Halo was developed in close collaboration with Qualcomm.

Announcements and coverage are on the [press page](https://scam.ai/resources/press).

### What we believe

Everyone deserves to know what's real. Detection should come with evidence, run where the risk is — on-device when privacy demands it — and evolve as fast as the attacks do.

### Research in the open

The detection work is published, not just claimed: 22 papers across 7 research programs — deepfake detection, document forgery, and age-estimation robustness among them — with open datasets and named authors on arXiv. That research feeds directly into the Eva V1.6 Detection Model, which is retrained as new generators appear.

### How we handle your data

Media sent to Scam AI is analyzed to produce a result and then discarded — we keep the results and the evidence, not your content. Compliance documentation, our DPA, and the subprocessor list are published in the trust center for security review.

## Careers
URL: https://scam.ai/company/careers

Don't apply if you just want a job — this isn't one. Everyone here shares the same goal and the same faith: beat the criminals using AI to deceive people. The bar is high and we are ruthlessly selective. But if this is your fight, we want to hear from you.

### This is not a regular job

We're a small team with one enemy: the people using generative AI to steal, impersonate, and defraud. Everyone is here for the mission, not the title — and for the refusal to let seeing stop meaning believing.

### A very high bar

We are selective on purpose. We hire people who are exceptional at what they do and who care more about the outcome than the org chart. You'll own real surface area from week one and ship work that stands directly between scammers and the people they target.

### Who we're looking for

We're hiring across Business Development and GTM, Marketing and Growth, and Detection Research and Engineering. If you can reach the teams under attack — or sharpen the detection that protects them — use the form at the end of this page to tell us what you'd take on and why this is your fight.

## Team
URL: https://scam.ai/company/team

The researchers and engineers building Scam AI — focused on detecting AI-generated deception at the moment it matters.

### Who we are

A small, senior team spanning detection research, ML engineering, and product, backed by operators and researchers from across security and AI.

### How we work

Close to the adversary and close to the customer: we retrain against new attack techniques as they appear and ship detection people can act on the same week.

# API documentation

## Start here
URL: https://scam.ai/docs/start-here

Run your first Scam AI detection in three steps: create an API key, send one file to POST /v1/detections, and read the verdict and score in the reply.

Covers: start here quickstart first request three steps get a key send a file read the verdict curl python node sdk @scam-ai/sdk scamai pip npm x-api-key POST /v1/detections base url https://api.scam.ai dashboard app.scam.ai api keys LIKELY_REAL ALERT LIKELY_AI score credits_used

## Which input do I send?
URL: https://scam.ai/docs/which-input

Photo, video, audio or a video link: which field to send, the accepted formats, the size cap, the credits charged and the evidence each one returns.

Covers: which input do I send image video audio link table field file url formats jpg png webp mp4 mov webm mp3 wav flac m4a aac ogg unsupported media type size cap 10 MB 100 MB 50 MB credits charged evidence frames segments source media type inferred

## Detect an image
URL: https://scam.ai/docs/detect-image

Send a photo or still image to POST /v1/detections as a file and read back one verdict, one score and the credits charged, with every response field explained.

Covers: detect an image photo still picture jpeg png webp ai-generated edited face swap send the file file parameters response fields id status created_at media model verdict score summary credits_used zero_charge_reason example 422 no face undecodable_image

## Detect a video
URL: https://scam.ai/docs/detect-video

Send a clip to POST /v1/detections as a file or a link and read back a verdict for the whole clip plus a per-frame timeline, with every field explained.

Covers: detect a video clip mp4 mov webm mkv avi deepfake face swap send the file send a link file url thumbnails thumbnail_b64 parameters response fields threshold_used frames index timestamp_s score cause timeout no_face example why 20 frames sampled one frame a second capped credits per frame

## Detect audio
URL: https://scam.ai/docs/detect-audio

Send a recording to POST /v1/detections as a file and read back a verdict for the whole recording plus a per-window timeline, with every field explained.

Covers: detect audio recording voice clone synthetic voice mp3 wav flac m4a aac ogg send the file file parameters response fields duration_ms segments start_ms end_ms value status scored abstained reason_codes no_voice_content window example credits per minute two verdicts

## Detect from a link
URL: https://scam.ai/docs/detect-link

Send a public video page or a direct video URL to POST /v1/detections and read back the clip's verdict plus a source block that says how much of it was scored.

Covers: detect from a link url public video page direct link to a video file paste link send a link thumbnails response fields source url canonical_url platform title uploader duration_s truncated analyzed_duration_s example what does truncated mean slivers 20 spread across long video 5 minutes too long invalid_url link_not_resolvable 501

## Reading AI detection results
URL: https://scam.ai/docs/reading-the-result

Learn how to interpret Scam AI verdicts, likelihood scores, per-frame evidence, null values and abstentions, then route results through your policy.

Covers: reading AI detection results verdict LIKELY_REAL ALERT LIKELY_AI routing signal score likelihood probability 0 to 1 threshold_used frames frame score cause timeout not_attempted segments duration_ms abstained reason_codes no_voice_content source truncated analyzed_duration_s null absent NOT_ASSESSABLE policy review step-up label hold independent evidence summary id status created_at media model Eva V1.6 credits_used zero_charge_reason

## Retrieve a detection
URL: https://scam.ai/docs/retrieve-detection

GET /v1/detections/{id} returns a detection you already ran, exactly as the create call answered it, so a later lookup needs no copy of your own.

Covers: retrieve detection GET /v1/detections/id fetch by id read back owner only stored response every run stored 404 NOT_FOUND history detections.get

## Authentication
URL: https://scam.ai/docs/authentication

Authenticate Scam AI API requests with the x-api-key header. Create, rotate and revoke keys yourself in the dashboard, and keep every key on the server side.

Covers: authentication x-api-key header api key create rotate revoke keys dashboard app.scam.ai/api-keys server-side secret never client-side bearer 401 sandbox key SCAMAI_API_KEY

## Errors
URL: https://scam.ai/docs/errors

Every error the Scam AI detection API returns, from 400 and 401 through 402, 413, 422, 429 and 503 to 504: the code each carries and the action that clears it.

Covers: errors 400 NO_FILE_UPLOADED INVALID_URL TOO_MANY_FILES UNEXPECTED_FILE_FIELD INVALID_UPLOAD 401 unauthorized api key 402 INSUFFICIENT_CREDITS 403 API_KEY_SCOPE FORBIDDEN EMAIL_VERIFICATION_REQUIRED 404 NOT_FOUND 413 FILE_TOO_LARGE LINK_MEDIA_TOO_LARGE 422 UNSUPPORTED_MEDIA_TYPE UNDECODABLE_IMAGE LINK_NOT_RESOLVABLE NOT_ASSESSABLE reasons no_face 429 too many requests rate limit 500 DETECTORS_UNAVAILABLE INTERNAL_ERROR 501 LINK_RESOLUTION_UNAVAILABLE 502 DETECTION_NOT_BILLABLE LINK_RESOLUTION_FAILED 503 BILLING_UNAVAILABLE 504 DETECTION_PLATFORM_TIMEOUT DETECTION_PLATFORM_UNREACHABLE LINK_RESOLUTION_TIMEOUT retryable resolverCode request_id support

## Credits and limits
URL: https://scam.ai/docs/credits-and-limits

What each media type costs in Scam AI credits, the size cap per media type, the rate limits, when a repeat file is not charged, and how sandbox keys behave.

Covers: credits and limits pricing cost per media type image 1 credit video per scored frame capped 20 audio per minute rounded up minimum 1 size caps image 10 MB video 100 MB audio 50 MB 413 rate limits 100 per minute per key 429 X-RateLimit RateLimit headers retry-after duplicate same file 30 days zero_charge_reason sandbox key balance dashboard

## SDKs, CLI and dashboard
URL: https://scam.ai/docs/sdks-and-dashboard

Install the TypeScript or Python SDK or the command-line client, manage keys, credits and detection history in the dashboard, and reach support with a run id.

Covers: sdks sdk client library typescript node javascript npm install @scam-ai/sdk python pip install scamai cli command line npx @scam-ai/cli check terminal official wrapper multipart typed errors exceptions retries never re-bill dashboard app.scam.ai usage api keys detection history support support@scam.ai on-prem vpc soc 2 gdpr trust center

## OpenAPI spec
URL: https://scam.ai/docs/openapi-spec

The OpenAPI 3.0 contract behind these docs, served at a stable URL for codegen, import into an API client, or a platform that consumes OpenAPI schemas.

Covers: openapi spec schema machine readable contract codegen generate client import swagger api.scam.ai/openapi.json api.scam.ai/openapi.yaml agent platform partner integration

## MCP server
URL: https://scam.ai/docs/mcp-server

Give an AI agent one tool that checks whether an image, a video or a recording is real: the hosted Scam AI MCP server, or run it yourself with npx.

Covers: mcp server model context protocol ai agent tool claude cursor vs code hosted mcp.scam.ai/mcp x-api-key run it yourself npx @scam-ai/mcp-server stdio SCAMAI_API_KEY node 22 mcp.json mcpServers detect_media get_credit_balance get_usage url file_path file_base64 local file

# Partners
URL: https://scam.ai/resources/blog/category/partners

## Qualcomm
Qualcomm develops computing, AI and connectivity technologies. Its Snapdragon and Dragonwing platforms serve mobile devices, PCs, automotive systems and connected products.
Website: https://www.qualcomm.com/

## ALCHERA
ALCHERA develops Vision AI for facial recognition, document OCR, identity verification and visual anomaly detection. It also provides custom AI data services.
Website: https://www.alchera.ai/

## TinyFish
TinyFish provides web infrastructure for AI agents. Its platform combines search, content extraction, browser sessions and web automation through APIs.
Website: https://www.tinyfish.ai/

## Modulate
Modulate develops audio AI models and APIs for conversational understanding, with tools for voice analysis, transcription, deepfake detection and content triage.
Website: https://www.modulate.ai/

## Fish Audio
Fish Audio provides AI voice tools for creators and developers, including text to speech, speech to text, voice cloning and conversational voice agents.
Website: https://fish.audio/

## Cozystay
Cozystay builds AI-powered operating software for alternative accommodations, connecting pricing, guest communication, task dispatch, quality assurance and settlement in one workflow.
Website: https://www.cozystay.com/

## deepidv
deepidv provides identity verification and fraud services through APIs and SDKs, including document checks, liveness detection, face matching, deepfake detection and compliance screening.
Website: https://www.deepidv.com/

## Eteon
Eteon develops an agentic AI ecosystem with reasoning, multimodal vision and predictive analytics agents for organisational knowledge, workforce management and decision support.
Website: https://www.eteonlabs.com/

## Aident AI
Aident builds tools that connect AI agents to apps and reusable skills. Its Loadout platform provides action visibility and credential storage.
Website: https://aident.ai/

## KrosAI
KrosAI provides phone infrastructure for AI agents, connecting them to local numbers and telephone networks with call routing, carrier integration and fraud prevention.
Website: https://krosai.com/

## Honestly
Honestly organizes opinions from social-media posts, videos and comments by product and feature, helping marketing, product and customer-experience teams understand feedback.
Website: https://www.usehonestly.com/

## Whitney AI
Whitney AI provides a workspace for model post-training. Teams can investigate agent failures, build evaluations, train models and monitor behavior in production.
Website: https://www.trywhitney.com/

## PovChat AI
PovChat AI is an AI roleplay and storytelling chatbot available on the web, iOS and Android. Create your own characters for companionship, RPG adventures and interactive stories, with message editing and image generation.
Website: https://povchat.ai/

# Site-wide FAQ
URL: https://scam.ai/faq

Q: What can Scam AI actually detect?
A: Scam AI detects deepfaked faces, face swaps, and fully AI-generated images and video. It also catches forged and tampered identity documents — including template-generated and synthetic IDs — manual edits and splices, and device-level attacks like virtual cameras and injection attacks that defeat naive selfie checks. One engine, the Eva V1.6 Detection Model, covers onboarding selfies, identity documents, claims media, live video meetings, and user-generated content.

Q: How accurate is it, and what about false positives?
A: Detection is probabilistic, not a yes/no: the Eva V1.6 Detection Model returns a tiered likelihood result — likely real, alert, or likely AI — with the score behind it and a plain-English summary of what it found. You set the thresholds — auto-decline, step-up, or human review — and tune them per flow: stricter for high-value wires, looser for routine account changes. You control the false-positive/false-negative trade-off, not a vendor's fixed cutoff. Eva V1.6 is retrained as new generators appear, so accuracy tracks the current attack landscape, not a frozen benchmark.

Q: Can it work in real time, including live calls?
A: Real-time detection is what Halo, Scam AI's meeting product, is built for. Halo runs deepfake detection on-device during live video meetings and flags a deepfaked face or manipulated feed while the call is happening, and because analysis stays on your machine, meeting video never leaves it. For application flows, the REST API offers real-time latency for live onboarding and step-up checks, plus batch endpoints for media libraries and claims backlogs. A face-swapped selfie gets declined before the account exists.

Q: Can detection run without sending media to Scam AI?
A: Yes. Halo analyzes meeting video on-device, so that media never leaves your endpoint at all. For data-residency or no-egress requirements, on-prem and private-cloud deployment keep media, results, and logs entirely inside your own perimeter.

Q: How do you keep up with new deepfake generators?
A: New face-swap and diffusion tools appear constantly, so Eva V1.6 trains on real attack media — actual fraud attempts, not academic datasets — and retrains as new generators show up in the wild. The team publishes original research — papers and open datasets on deepfake detection, document forgery, and adversarial attacks on age-estimation systems, with named authors on arXiv — and studied techniques become detection signals. Ask any vendor when its model was last updated, and against what.

Q: How does it integrate with our stack?
A: Integration is one REST API: send video, images, or audio; get back a verdict, the score behind it, and a plain-English summary of what was found. Drop it into your existing decision point — onboarding/KYC, claims intake, the support or moderation queue — and route on the score: approve, step-up, or review. There is no model to train, so a first integration typically takes a day, not a quarter. Node and Python SDKs and a command-line tool wrap the API; Halo, a lightweight desktop app, needs no integration; regulated environments can deploy on-prem or private-cloud. Document forgery checks are a separate plan, arranged with our team.

Q: Are you compliant — SOC 2, GDPR, data residency?
A: Scam AI is SOC 2 Type II audited and GDPR compliant, with a Data Processing Agreement available. The legal stack — privacy policy, terms, DPA, MSA, subprocessor list, and the Halo EULA — is published openly on the site. Banks and public-sector agencies with data-residency or no-egress requirements can deploy on-prem or private-cloud: the same models run inside your perimeter, so nothing crosses a boundary you don't control. Halo analyzes meeting video on-device, so meeting content never reaches Scam AI's infrastructure.

Q: What evidence do we get for a result?
A: Every result ships with a tiered verdict — likely real, alert, or likely AI — the score behind it, and a plain-English summary of what the model found; video and audio also carry the per-frame and per-window series behind the score. Analysts get something to review instead of a bare number, and auditors can trace why an application was declined or escalated. Named per-signal breakdowns and examiner-grade forensic reports are arranged with our team — ask us on a demo call.

Q: How is this different from a single-point detector?
A: Single-point detectors answer one question — "does this media look AI-generated?" — and stop. Scam AI layers multimodal detection on top of your existing KYC/IDV vendor, strengthening it, not replacing it, and adds document-forgery detection, device signals, and on-device meeting protection in one combined result. A flagged selfie is only useful tied to the identity being onboarded and an approve/decline/review decision; a point detector leaves that connective work to your engineers. With a platform, flag, identity, and decision arrive together.

Q: What does it cost?
A: Self-serve is one flat rate: $0.02 a credit, prepaid, and credits never expire. An image check is 1 credit, video is 1 credit per sampled frame (capped at 20 a clip), audio is 1 credit per minute — the full table is on /pricing. Document detection and enterprise volume are priced on request: bring your use case and monthly volumes to a 15-minute demo and leave with a price mapped to your actual flow.

# Pricing and limits
URL: https://scam.ai/pricing

Credit cost per run: image 1 credit per image; video up to 20 credits per clip (1 credit per scored frame, about one frame a second, capped at 20; a frame with no score is not charged); audio 1 credit per minute, rounded up, minimum 1. A credit is $0.02.
Upload limits enforced by the scam.ai website and dashboard: image 10 MB, video 25 MB, audio 25 MB. The REST API itself accepts image 10 MB, video 100 MB, audio 50 MB; anything that is not a photo, video or audio file is refused with 422 UNSUPPORTED_MEDIA_TYPE (https://scam.ai/docs/credits-and-limits).
Self-serve pricing is prepaid credits at one flat rate: add credits first, then every run spends them. Enterprise plans (a monthly platform fee with a lower per-frame rate) are arranged through https://scam.ai/contact; details at https://scam.ai/pricing.

# Site map

## Start here
- [Home](https://scam.ai/): Deepfake detection and scam prevention.
- [Pricing](https://scam.ai/pricing): Pay as you go at $0.02 a credit, one flat rate; enterprise plans and document detection through sales.
- [Why Scam AI](https://scam.ai/why-scamai): Detection you can act on, with evidence.
- [Technology](https://scam.ai/technology): The detection research behind Scam AI.
- [FAQ](https://scam.ai/faq): Direct answers on what Scam AI detects, accuracy, real-time use, data handling, and compliance.
- [Request a demo](https://scam.ai/contact)

## Platform
- [Platform overview](https://scam.ai/platform): Detection, analytics, and the API in one place.
- [Eva V1.6 Detection Model — our detection engine](https://scam.ai/platform/eva): The unified API behind Scam AI's detection results: video, images, and audio, scored in real time.
- [Check if AI — spot AI-generated media in seconds](https://scam.ai/platform/check-if-ai): Upload an image and get an instant read on whether it's AI-generated.
- [Deepfake Playground](https://scam.ai/platform/deepfake-playground): Deepfake yourself in 30 seconds. Point your camera at your own face, pick someone to become, and watch the swap happen live — the same thing an attacker can do with consumer tools.
- [On-prem deepfake detection](https://scam.ai/platform/on-prem-deployment): Run Scam AI detection inside your own environment — VPC, private cloud, or fully on-premises — so regulated data never leaves your perimeter.
- [PII Zero: Claims Media De-identification](https://scam.ai/platform/pii-zero): Offline redaction of policyholder identity in claim documents, loss photos and video: names, addresses, account and policy numbers, faces and plates.
- [Risk Signals](https://scam.ai/platform/risk-signals): Turn detection results into a view of your threat surface — patterns rather than single incidents. Available on request.
- [Deepfake Detection](https://scam.ai/platform/deepfake-detection): Purpose-built deepfake signals — face swaps, full generations, and replays — scored across images, video, and live calls.
- [ID Document Forgery Detection](https://scam.ai/platform/id-document-forgery): Tampering and forgery signals for identity documents — passports, national IDs, driver's licenses, residence permits.
- [Loan & Insurance Document Forgery](https://scam.ai/platform/loan-insurance-forgery): Forgery signals for due-diligence documents — bank statements, pay stubs, proof of income and address, invoices, policy documents.
- [Age Estimation](https://scam.ai/platform/age-estimation): Age checks from a live selfie — an estimated age and the probability the person is over your threshold, no ID collected.
- [Device Signals — liveness & presentation attacks](https://scam.ai/platform/capture-signals): Detection for spoofed presentations — printed photos, screen replays, and masks — plus liveness failures, caught at the point of capture.
- [Manual-Edit Forensics — image forgery detection](https://scam.ai/platform/manual-edit-forensics): Pixel-level forensics for manually edited selfies and documents — splices, clones, and retouching.
- [Verifit | Remote inspections. Confident decisions.](https://scam.ai/verifit): Cut inspection costs and keep decisions moving. Verify property and assets remotely, with fewer site visits and less back-and-forth.
- [How Verifit works | Less chasing. Faster answers.](https://scam.ai/verifit/how-it-works): Make inspections easier with Verifit. Send a link, collect verified photos, and give your team a faster way to review cases with fewer customer follow-ups.
- [Verifit for insurance | Fast claims, fewer visits](https://scam.ai/verifit/insurance): Move claims forward and reduce inspection costs with Verifit. Verify property remotely, simplify policyholder follow-ups, and flag suspicious submissions.
- [Verifit for lending | Keep funding moving](https://scam.ai/verifit/lending): Keep inspections from holding up funding. Verify property, progress, and equipment remotely with Verifit to reduce site visits and move decisions forward.
- [See Verifit in action | Request a demo](https://scam.ai/verifit/demo): See how Verifit can help reduce inspection costs and speed up decisions. Request a demo tailored to your workflow, your team, and your customer experience.
- [Halo — on-device meeting toolkit](https://scam.ai/halo): On-device meeting notes, speech-to-text, and real-time deepfake detection for Windows.

## For developers and AI agents
- [MCP & Skills for LLMs](https://scam.ai/platform/mcp): Give an AI agent one tool that checks whether an image, a video or an audio recording is real, and a Skill that tells it what to do with the answer.
- [Detection API](https://scam.ai/platform/api): One API for the places fraud happens: send an image, a video or a recording and get a scored verdict back. SDKs for Node and Python, a CLI and an MCP server are published.
- [API documentation](https://scam.ai/docs): Run your first Scam AI detection in three steps: create an API key, send one file to POST /v1/detections, and read the verdict and score in the reply.
- [API docs: Which input do I send?](https://scam.ai/docs/which-input): Photo, video, audio or a video link: which field to send, the accepted formats, the size cap, the credits charged and the evidence each one returns.
- [API docs: Detect an image](https://scam.ai/docs/detect-image): Send a photo or still image to POST /v1/detections as a file and read back one verdict, one score and the credits charged, with every response field explained.
- [API docs: Detect a video](https://scam.ai/docs/detect-video): Send a clip to POST /v1/detections as a file or a link and read back a verdict for the whole clip plus a per-frame timeline, with every field explained.
- [API docs: Detect audio](https://scam.ai/docs/detect-audio): Send a recording to POST /v1/detections as a file and read back a verdict for the whole recording plus a per-window timeline, with every field explained.
- [API docs: Detect from a link](https://scam.ai/docs/detect-link): Send a public video page or a direct video URL to POST /v1/detections and read back the clip's verdict plus a source block that says how much of it was scored.
- [API docs: Reading AI detection results](https://scam.ai/docs/reading-the-result): Learn how to interpret Scam AI verdicts, likelihood scores, per-frame evidence, null values and abstentions, then route results through your policy.
- [API docs: Retrieve a detection](https://scam.ai/docs/retrieve-detection): GET /v1/detections/{id} returns a detection you already ran, exactly as the create call answered it, so a later lookup needs no copy of your own.
- [API docs: Authentication](https://scam.ai/docs/authentication): Authenticate Scam AI API requests with the x-api-key header. Create, rotate and revoke keys yourself in the dashboard, and keep every key on the server side.
- [API docs: Errors](https://scam.ai/docs/errors): Every error the Scam AI detection API returns, from 400 and 401 through 402, 413, 422, 429 and 503 to 504: the code each carries and the action that clears it.
- [API docs: Credits and limits](https://scam.ai/docs/credits-and-limits): What each media type costs in Scam AI credits, the size cap per media type, the rate limits, when a repeat file is not charged, and how sandbox keys behave.
- [API docs: SDKs, CLI and dashboard](https://scam.ai/docs/sdks-and-dashboard): Install the TypeScript or Python SDK or the command-line client, manage keys, credits and detection history in the dashboard, and reach support with a run id.
- [API docs: OpenAPI spec](https://scam.ai/docs/openapi-spec): The OpenAPI 3.0 contract behind these docs, served at a stable URL for codegen, import into an API client, or a platform that consumes OpenAPI schemas.
- [API docs: MCP server](https://scam.ai/docs/mcp-server): Give an AI agent one tool that checks whether an image, a video or a recording is real: the hosted Scam AI MCP server, or run it yourself with npx.

## Research
- [Research](https://scam.ai/research): Published papers on deepfake detection, document forgery, and age-estimation robustness, with open datasets and named authors on arXiv.

## Solutions
- [Solutions overview](https://scam.ai/solutions)
- [KYC / IDV fraud detection](https://scam.ai/solutions/kyc-idv): Make sure the person enrolling is live, present, and real — not a replay, a render, or a synthetic identity. Scam AI works with your KYC vendor to strengthen it, not replace it.
- [Loan & credit fraud detection](https://scam.ai/solutions/loan-credit): Stop synthetic applicants and doctored income documents at origination, before the money moves.
- [Insurance claims fraud detection](https://scam.ai/solutions/insurance-claims): Verify claim photos, videos, and documents are genuine — not staged, edited, or AI-fabricated.
- [Meeting deepfake protection](https://scam.ai/solutions/meeting-security): Continuous verification of who is really on the call — powered by Halo.
- [Pig butchering scams](https://scam.ai/solutions/pig-butchering): Detect the fake faces and forged proof behind long-con romance and investment scams — before your users are drained.
- [Deepfake social engineering attacks](https://scam.ai/solutions/social-engineering): Catch the deepfaked faces attackers use to manipulate employees into transfers, access, and approvals.
- [HR / Hiring](https://scam.ai/solutions/hiring-fraud): Protect hiring end to end — make sure the person you interview is the person you hire, not a deepfake stand-in or a proxy.
- [Content moderation](https://scam.ai/solutions/content-moderation): Score uploads for AI generation and manipulation before synthetic content spreads.
- [Misinformation detection](https://scam.ai/solutions/misinformation): Detect the deepfakes and synthetic media behind misinformation before it spreads and erodes trust.
- [Deepfake brand protection](https://scam.ai/solutions/brand-protection): Catch the deepfakes, fake ads, and impersonations that hijack your brand and your executives' likenesses.
- [AI phishing & spam detection](https://scam.ai/solutions/phishing-spam): Screen the images, documents, and media inside phishing and spam campaigns for AI generation and manipulation.

## Industries
- [KYC/AML compliance](https://scam.ai/industries/kyc-aml-compliance): Meet KYC and AML obligations against AI-era fraud — deepfake selfies, forged IDs, and synthetic identities caught at onboarding and beyond.
- [On-device deepfake detection agent](https://scam.ai/industries/on-device-agent): Real-time detection that runs on the device — media and identity verified locally, with nothing leaving the endpoint. Shipping today in Halo; an embeddable agent for your own apps and devices is in development.
- [Deepfake detection for financial services](https://scam.ai/industries/financial-services): Stop synthetic identities at onboarding and deepfaked executives on video calls before the wire is released.
- [Deepfake detection for the public sector](https://scam.ai/industries/public-sector): Protect officials and citizens from impersonation campaigns.
- [E-commerce & marketplaces](https://scam.ai/industries/e-commerce-marketplaces): Keep fake sellers and cloned storefronts out.
- [Dating apps](https://scam.ai/industries/dating-apps): Keep fake profiles, AI-generated photos, and deepfaked video calls off your platform — before they turn into romance scams.
- [Telecom fraud detection](https://scam.ai/industries/telecom): Protect SIM changes, account access, and support channels from synthetic identities and social engineering.

## Compare


## Resources
- [Resources](https://scam.ai/resources): Research, documentation, the regulation tracker, and partner programs.
- [Deepfake regulation tracker](https://scam.ai/resources/regulation): Where deepfake and AI-fraud law stands, jurisdiction by jurisdiction.
- [Education](https://scam.ai/resources/education): How AI-driven scams work, and how to check for them: tools you can try on your own media, published research, and the rules that apply.
- [Trust Center](https://scam.ai/resources/trust-center): Security, privacy, and compliance documentation.
- [Changelog](https://scam.ai/resources/changelog): What's new in Scam AI detection — models, APIs, and product.
- [Affiliate program](https://scam.ai/resources/affiliate): Earn by referring teams that need deepfake and fraud detection.
- [Partner program](https://scam.ai/resources/partners): Resell, integrate, or build on the Scam AI detection API.
- [Deepfake detection education program](https://scam.ai/resources/education-program): Training, certification, and curriculum for partners, students, and institutions who want to teach and deploy deepfake detection.
- [Partners: companies we work with](https://scam.ai/resources/blog/category/partners): Meet our partners, visit their websites and read our collaboration articles.

## Blog
- [Blog](https://scam.ai/resources/blog): Detection research and field notes from the team.
- [Deepfake detection accuracy: how to interpret the evidence](https://scam.ai/resources/blog/how-to-interpret-ai-detection-evidence): Learn how to read detector scores, verdicts, localized evidence and abstentions, then route media to approval, step-up checks or human review.
- [Document fraud detection software: choosing it for lending](https://scam.ai/resources/blog/choosing-document-fraud-detection-software): Compare document fraud detection software by evidence coverage, reviewable results, pilot design, deployment and cost. A buying guide for lenders and fintechs.

## News
- [Press](https://scam.ai/resources/press): Announcements and coverage.

## Community
- [Anti-Scam Community](https://scam.ai/community): Threat sharing, researcher Q&A, and early access.

## Company
- [About Scam AI](https://scam.ai/company/about): Scam AI exists because seeing is no longer believing. We build the tools that let people and organizations trust what they see.
- [Careers](https://scam.ai/company/careers): Don't apply if you just want a job — this isn't one. Everyone here shares the same goal and the same faith: beat the criminals using AI to deceive people. The bar is high and we are ruthlessly selective. But if this is your fight, we want to hear from you.
- [Team](https://scam.ai/company/team): The researchers and engineers building Scam AI — focused on detecting AI-generated deception at the moment it matters.
- [Company](https://scam.ai/company): About, team, careers, and brand.
- [Brand kit](https://scam.ai/brand)

## Legal
- [Privacy Policy](https://scam.ai/legal/privacy): How Reality Inc. (DBA Scam.ai) collects, uses, discloses, and protects personal data across its sites and services, and the rights available to individuals.
- [Terms of Service](https://scam.ai/legal/terms): The binding terms governing access to and use of the ScamAI website, dashboard, API, and apps, including accounts, acceptable use, termination, and liability.
- [Cookie Policy](https://scam.ai/legal/cookies): How and why ScamAI uses cookies and similar technologies on its sites, what each category does, and how to control or withdraw consent in your browser or ours.
- [Data Processing Addendum](https://scam.ai/legal/dpa): Processor terms for personal data ScamAI processes on behalf of customers under GDPR Article 28: instructions, security, subprocessors, transfers, audits.
- [Subprocessors](https://scam.ai/legal/subprocessors): Third parties ScamAI engages to help operate the Services — hosting, database, email, analytics, and abuse prevention — and how we announce changes to the list.
- [Master Services Agreement](https://scam.ai/legal/msa): The default commercial terms for paid ScamAI Services — fees, term, confidentiality, IP and data, and liability — unless a signed agreement provides otherwise.
- [Halo App End User License Agreement](https://scam.ai/legal/halo-eula): The license terms for the Halo on-device deepfake-detection app: what you may install and run, what stays on your device, updates, and limits of liability.


# Blog

## Deepfake detection accuracy: how to interpret the evidence
URL: https://scam.ai/resources/blog/how-to-interpret-ai-detection-evidence
Category: Insights
Published: 2026-10-07T00:00:00.000Z

Deepfake detectors turn complex forensic signals into a score, and that simplicity invites overreach. A business may read a score as certainty, treat a low-risk verdict as proof, or send every inconclusive result to the same queue. Each move strips context from the evidence. A safer decision accounts for the result, the media and the consequences of being wrong. At Scam AI, we return the layers of evidence needed for that process.

## At a glance

- A score is a likelihood signal for one input. It is not the detector's accuracy or proof of fraud.
- A verdict selects a policy path. Your organization owns the action at the end of that path.
- Localized evidence tells a reviewer where to inspect. It does not make the highlighted area ground truth.
- A null value or abstention is a gap in the evidence, never a favorable result.
- High-impact decisions need representative testing, corroboration and accountable review.

## Deepfake detection accuracy starts with the test conditions

Accuracy is a property of an evaluation. It is not a property of the score attached to one file.

A reported accuracy figure only becomes useful when its methodology names:

- the image, video or audio dataset and how ground truth was established
- the mix of genuine and manipulated media
- the manipulation methods and generation tools represented
- benign changes such as compression, cropping, resizing, background noise and screen recording
- the model version and decision rule used for the test
- how unscored and abstained cases were counted

Without that context, two accuracy figures may describe very different tasks. A detector tested on clean, balanced research data faces a different problem from a detector receiving recompressed uploads, short voice notes and unfamiliar generation methods.

The [AI Risk Management Framework](https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/) from the National Institute of Standards and Technology (NIST) calls for accuracy measurements to use realistic test sets that represent expected conditions, with the methodology documented. NIST's [OpenMFC evaluations](https://mfc.nist.gov/) also separate file-level detection from manipulation localization and define the operating point used to assess false alarms. Those are useful design principles for a buyer evaluation, even when the production workflow uses different measures.

This is why a single headline number cannot answer the operational question: what should we do with this file?

## Read a detection result as a stack of evidence

A useful result has several layers. Each layer answers a different question.

| Evidence layer | What it tells you | What it cannot establish |
| --- | --- | --- |
| Score | How strongly the model's learned signals align with AI generation or manipulation patterns for this input | Overall detector accuracy, the sender's intent or proof of fraud |
| Verdict | The product's condensed interpretation of the result | Your organization's final business decision |
| Localized evidence | Which frame, audio window or image region deserves inspection, when supported | That every other part is clean or that the highlighted area is ground truth |
| Missing evidence or abstention | Where the system could not produce a valid result and why | A clean result |
| Provenance and capture evidence | What is known about the file's source, history or capture path | Whether the depicted event is true |
| Case context | Whether the media is consistent with known channels, records and behavior | A forensic assessment on its own |

The layers remain distinct, even though they feed the same policy route and human decision.

![Five layers of detection evidence feeding a policy route and human decision](/blog/how-to-interpret-ai-detection-evidence/detection-evidence-stack-37fc80b7db79.png)

Our result structure exposes these layers directly. The [result documentation](https://scam.ai/docs/reading-the-result) defines `verdict`, `score` and a human-readable `summary`. Video responses add per-frame evidence and a cause beside an unscored frame. Audio responses add per-window values, an abstention status and reason codes. The response also records the model, detection ID and completion timestamp for traceability.

### Score: a likelihood signal for this input

The score is model evidence. It can help order a review queue, compare cases inside the same tested workflow and study where a routing policy should place more scrutiny.

It does not tell you the percentage of all deepfakes the detector catches. It should not be translated into a statement such as “this customer is that percent likely to be committing fraud.” Interpreting a score as a real-world probability requires a documented calibration study on representative deployment data. Keep it as a likelihood signal when that evidence is unavailable.

Our video results can also include `threshold_used`. This is the line used to interpret per-frame scores for that run. It is separate from the business rules that determine approval, review or decline.

### Verdict: a compact model interpretation

We use three verdict labels across Scam AI media results:

- `LIKELY_REAL` means the detector found no strong signs of AI generation or manipulation. It does not authenticate the person, source or event.
- `ALERT` means AI-related signals appeared and the result is inconclusive. This belongs in a review or step-up path.
- `LIKELY_AI` means the detector found strong signs of AI generation or manipulation. Consequential action still follows the customer's policy and supporting evidence.

Use a verdict to select a policy path. The customer remains responsible for the decision at the end of that path.

### Localized evidence: where to look

A whole-file score compresses time and space. Local evidence restores some of that detail.

For video, a reviewer can inspect the frames and timestamps that produced evidence, then watch the surrounding sequence. For audio, the reviewer can listen to a flagged window with the speech immediately before and after it. An image detector may return a region or heatmap when that capability is supported.

Localization makes a result easier to investigate, but it does not turn the highlighted area into proof. A model may respond to compression, compositing or capture artifacts that overlap with synthetic-media patterns. Reviewers need the original file, the local evidence and the case context together.

The documented example below shows Scam AI's result schema rather than a production detection. It illustrates how the response explains a missing frame score.

![Scam AI result documentation showing an unscored frame cause](/blog/how-to-interpret-ai-detection-evidence/result-evidence-documentation-cf8337616dcf.png)

### Missing evidence and abstention: an explicit gap

An abstention is a valid system state. It means the detector did not have enough usable evidence to score all or part of the input.

In our result structure, a video frame can carry a null score with a cause such as a timeout. An audio window can be marked `abstained` with reason codes such as no voice content. Null does not mean zero, genuine or low risk.

Route that gap according to its cause. A corrupted or unsupported input may need a new capture in a supported format. A missing speech segment may require another recording. A high-impact case with incomplete evidence belongs in review. Do not silently replace an abstention with a favorable score.

## Map the evidence to an action

The detector supplies evidence. Your policy assigns the action. A practical routing matrix looks like this:

| Evidence state | Candidate action | Required safeguard |
| --- | --- | --- |
| Complete result, no strong synthetic-media signals, corroborating checks are consistent | Approve when policy permits | Record that the detector found no strong signals; do not label the media “proven authentic” |
| Inconclusive result, conflicting checks or unusual case context | Step up | Request a fresh capture, use a known-channel callback, or obtain another independent record |
| Strong or localized synthetic-media signals in a consequential case | Human review | Preserve the original, inspect the localized evidence and document the reviewer rationale |
| Strong signals plus corroboration, within a workflow validated for this input type | Decline when the organization's policy permits | Apply appeal, exception and audit procedures appropriate to the decision |
| Missing result, abstention or unsupported input | Step up or review | Resolve the evidence gap; never treat missing analysis as a favorable result |

Result evidence is one input into the route. Corroboration and the consequence of a wrong outcome also belong in the customer-owned policy.

![Result evidence, corroboration and consequence feeding a customer policy with four review paths](/blog/how-to-interpret-ai-detection-evidence/evidence-to-action-routing-3cd3c47a7b99.png)

This matrix should vary by consequence. A content-moderation label, a request for recapture and the denial of a financial service do not carry the same harm if the detector is wrong.

A false alert sends genuine media into friction or review. A missed detection allows manipulated media to proceed. Lowering one type of error can raise the other, so there is no universal routing line. Choose a policy through representative testing, review capacity and the cost of each error in the actual workflow.

## Corroborate high-impact results

Independent checks answer questions that a detector score cannot.

**Provenance.** Content Credentials can record signed information about a file's origin and edit history. The [C2PA explainer](https://c2pa.org/specifications/specifications/2.2/explainer/Explainer.html) from the Coalition for Content Provenance and Authenticity (C2PA) is explicit that valid provenance does not establish whether the depicted content is true or factual. Missing credentials are also inconclusive because many legitimate capture and editing tools do not add them.

**Capture integrity.** Device attestation, virtual-camera checks, liveness controls and a fresh in-app capture can show whether media traveled through the expected path. These controls address delivery and presentation attacks. They answer a different question from content analysis.

**Known-channel verification.** A callback to an established number, confirmation inside an existing account session or direct contact with the purported sender can test the identity and request behind the media.

**Case records.** Transaction history, timestamps, prior submissions and source documents can reveal contradictions or support a benign explanation.

Run these checks as complementary evidence, rather than as a vote among similar scores. Do not assume multiple scores are independent when detectors use overlapping training data or signals. NIST's [synthetic-content report](https://www.nist.gov/publications/reducing-risks-posed-synthetic-content-overview-technical-approaches-digital-content) likewise treats provenance tracking and content-based detection as distinct approaches within a broader transparency system.

## Validate the workflow with a representative shadow pilot

Vendor benchmark results cannot replace testing on the media your organization receives. Use a representative shadow-mode pilot, where detection runs beside the current process without changing live outcomes.

1. **Define the decision and the harms.** Name the action the signal may influence, who can be affected and what happens after an incorrect route.
2. **Build a labeled sample.** Include genuine and known manipulated media across the real mix of modalities, sources, capture devices, quality levels and benign edits. Keep a documented labeling and adjudication process.
3. **Hold out evaluation cases.** Keep part of the labeled set separate from policy design so the final check measures the chosen workflow on unseen cases.
4. **Run blind and log every state.** Record the model version, score, verdict, local evidence, abstentions, processing failures and current human outcome without exposing the label to the detector or reviewer.
5. **Study errors by slice.** Review false alerts and missed detections separately by modality, source, quality and attack type. Count abstentions as their own outcome. A single aggregate accuracy figure can hide the route that creates operational harm.
6. **Set customer-owned actions.** Assign approve, step-up, review and decline paths based on the pilot evidence, decision severity and available review capacity. Validate the complete policy on the holdout set.
7. **Monitor after launch.** Track reviewer overrides, appeals, evidence gaps and shifts in input quality or attack methods. Re-run the evaluation when the model or workflow changes.

For remote identity proofing, [NIST SP 800-63A](https://pages.nist.gov/800-63-4/sp800-63a.html) provides a useful example of this method. It calls for testing automated image analysis against genuine and manipulated media, documenting the attack artifacts tested and augmenting automated decisioning with manual review to address errors.

## Keep a result record that another reviewer can reconstruct

A decision becomes defensible when another qualified person can understand what happened without guessing. Retain the fields your policy and privacy requirements allow:

- source or capture method and a stable identifier for the original media
- detection ID, model version and completion timestamp
- score, verdict and any result-level interpretation line
- localized evidence, missing values, causes and abstention reasons
- provenance, capture-integrity and known-channel checks
- action taken, reviewer identity and rationale
- recapture, exception and appeal outcomes

Retention should match the sensitivity of the media and the purpose of the review. Our [Scam AI Trust Center](https://scam.ai/resources/trust-center) explains our current security, privacy and media-handling approach. Your audit record should also capture the customer-owned policy version that produced the action.

Deepfake detection accuracy matters, but operational reliability comes from the whole evidence path: representative evaluation, transparent result fields, explicit gaps, corroboration and accountable review.

[Book a Scam AI demo](https://scam.ai/contact) to see how scores, verdicts and abstentions can support your review policy.

## Document fraud detection software: choosing it for lending
URL: https://scam.ai/resources/blog/choosing-document-fraud-detection-software
Category: Insights
Published: 2026-10-15T12:00:00.000Z

A bank statement can look convincing while containing an altered balance. A genuine statement can also raise an alert after an ordinary scan or conversion. For lenders and fintechs, both mistakes matter: one exposes the business to risk, the other adds friction for a legitimate applicant.

Choosing software means understanding what it can establish from an uploaded file, what evidence reaches an underwriter, and how uncertain results affect the application. The buying decision starts with the gap in your existing controls.

## What document fraud detection software establishes

Document fraud detection software analyzes submitted files for signs of alteration, fabrication or suspicious reuse. Depending on the product, it can examine file structure, visual inconsistencies, generation artifacts, internal arithmetic and relationships between submissions. The output supports a review decision.

Several questions remain separate:

- Has this file been manipulated?
- Do its amounts and dates agree with one another?
- Did the named institution issue it, and do its contents match that institution’s records?
- Does it support the applicant’s income, assets or eligibility under your lending policy?

A system can answer one question well while leaving the others open. Optical character recognition (OCR) extracts text. Correctly reading a balance does not establish that the balance is genuine. Reconciled arithmetic establishes internal consistency; a fabricated statement can still add up.

Generative AI also belongs in the threat model. The Financial Crimes Enforcement Network (FinCEN) has documented schemes involving altered or fabricated identity documents used to bypass financial institutions’ controls. Its [deepfake fraud alert](https://www.fincen.gov/system/files/shared/FinCEN-Alert-DeepFakes-Alert508FINAL.pdf) treats suspicious indicators as grounds for further scrutiny, rather than conclusive evidence.

At Scam AI, we recommend an authenticity layer when the unresolved question is whether uploaded paperwork has been altered or generated. Our [financial document checks](https://scam.ai/platform/loan-insurance-forgery) cover bank statements, pay stubs, proof-of-income and address documents, invoices and policy paperwork. Eva V1.6 is our detection model. We return scores and localized evidence for review, and surface suspicious reuse across applications. These checks sit alongside your financial and identity controls.

## Choose the control that supplies the missing evidence

For an upload-integrity gap, we recommend Scam AI’s document-authenticity layer. Other control categories serve different jobs. An extraction platform may be useful for cash-flow analysis; a document-authenticity tool may be useful for suspicious edits. Neither capability implies the other.

| Control category | The evidence it supplies | The boundary to preserve |
| --- | --- | --- |
| Document-authenticity analysis (Scam AI) | Signals of tampering, generation or suspicious reuse, with supporting detail where available | Does not by itself confirm account ownership, income or intent |
| OCR and document processing | Extracted fields, transaction rows and calculations | Accurate extraction can faithfully reproduce false information |
| Identity verification | Evidence connecting an applicant to an identity through the supported identity workflow | A verified person can submit a manipulated financial document |
| Source-based financial verification | Authorized information obtained from a bank, payroll provider or other relevant source | Coverage, authorization and the particular data returned determine what it establishes |
| Manual investigation | Case context, explanations, contradictions and independent corroboration | A visual review alone has limited access to hidden file structure |

Source-based verification is a distinct control. For example, Fannie Mae’s [asset verification guidance](https://selling-guide.fanniemae.com/sel/b3-4.2-01/verification-deposits-and-assets) includes direct depository-institution responses and authorized third-party verification alongside statement documentation. A detector score does not replace those source records or the lender’s applicable documentation requirements.

The practical buying brief names the unresolved decision. “Route suspicious income-document uploads to an underwriter before funding” is a workable scope. “Prevent loan fraud” leaves too much undefined to judge a proposal.

## Map coverage to the documents that reach origination

“Supports PDFs” is only the beginning of a coverage answer. A bank-downloaded PDF, a scanned printout and a phone photograph carry different evidence. The coverage description needs to separate document families from file formats, then explain which checks remain available for each combination.

The proposal needs to distinguish structural analysis from pixel analysis. Metadata and embedded text can provide clues in a digital file. A photograph of a printed statement no longer contains the original PDF’s structure. Conversely, an image-only check does not automatically cover every manipulation of an embedded text layer.

![Original PDFs supply file structure and page images; photographs supply page images.](/blog/choosing-document-fraud-detection-software/document-evidence-paths-d574f5eaa9d5.png)

Each signal also has a benign explanation. A changed template may be an issuer redesign. A font difference may come from a legitimate export. Metadata describes part of a file’s history and can be edited; the name of an editing application alone does not prove deception.

For financial documents, the coverage record should specify:

- The document families, issuers, languages and layouts in scope.
- Accepted input paths, including original PDFs, scans and photographs where supported.
- The checks performed on each path, including any page or file limits.
- Unsupported, unreadable and partially analyzed states, with their reasons.

A claim of AI-generation detection needs its own evidence. Traditional text-tampering research does not necessarily cover AI edits. The authors of the [DocTamper research dataset](https://github.com/qcf-568/DocTamper) explicitly state that it excludes AI-generated text tampering. Our buying principle is therefore to treat coverage of conventional edits and coverage of generative edits as separate evaluation questions.

## An underwriter needs more than a score

A score helps prioritize attention. It is not overall model accuracy or proof of fraud. A real-world fraud probability needs documented calibration on representative lending data; a score does not establish an applicant’s intent. A localized signal identifies an area for review; it does not establish ground truth for that region or certify the rest of the page.

A reviewable result makes the underlying question specific. An illustrative message such as “Potential alteration around the closing-balance field” gives an underwriter a place to begin. An unexplained risk label leaves the reviewer to reconstruct the concern.

The result record should contain the original-file identifier, pages analyzed, relevant signals and locations, model version, processing status and the time of analysis. The decision record adds the lender-owned routing policy, reviewer rationale and any corroborating evidence.

A successful processing response also needs to mean more than “the request completed.” Partial analysis, an unsupported document and an unreadable page need explicit states. An empty result cannot silently become a favorable result.

Our guide to [interpreting detection evidence](https://scam.ai/resources/blog/how-to-interpret-ai-detection-evidence) explains how scores, local signals and missing evidence remain distinct from business decisions. For a lender, that distinction needs to survive all the way into the origination system.

## A pilot should resemble the intake queue

A useful pilot runs beside the existing review process before its results change applicant outcomes. It includes genuine documents as well as confirmed manipulations, and it preserves how each label was established. A previous decline is not automatically a fraud label; an approved loan is not automatically proof that every submitted file was authentic.

The National Institute of Standards and Technology’s [AI risk framework](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/) calls for documented evaluation methods and performance demonstrated under conditions similar to deployment. Applied to lending intake, that means representative files and documented uncertainty, rather than a demonstration built only around obvious forgeries.

The following are illustrative evaluation cases, not customer outcomes or product benchmarks.

| Evaluation case | The buying question it answers |
| --- | --- |
| Genuine issuer-produced statement in its original format | Can the system process the ordinary intake path without unnecessary alerts? |
| Genuine document after a permitted scan, conversion or redaction | Does the result distinguish benign processing from suspected manipulation? |
| Confirmed alteration to an amount or applicant field | Does the evidence identify the relevant concern, rather than only flag the file? |
| Synthetic document with internally consistent figures | Does generation coverage extend beyond broken arithmetic? |
| Similar documents from unrelated legitimate applicants | Does reuse analysis account for the shared layouts issuers normally produce? |
| Multi-page submission with an unreadable or unsupported page | Is incomplete coverage visible to the reviewer and routing system? |

Synthetic examples need separate labels from real confirmed cases. A controlled edit supplies knowledge of the changed region, but it does not establish the sender’s intent or represent every attack a lender will encounter. Cases without a defensible label remain uncertain.

The pilot report should separate missed known manipulations, alerts on genuine files, unsupported inputs and processing failures. It should also show how much of the submitted material received analysis. Excluding unscored files can make the evaluated subset look reassuring while leaving a gap in the live queue.

Headline accuracy alone cannot express that gap. Relevant measures include detection of known manipulations, false alerts on genuine documents, analysis coverage and the review workload created at the chosen routing threshold. No single measure settles the purchase.

A separate set of cases, held back from threshold selection, gives the final policy evaluation a less biased basis. Results also need separation by document type and intake path, so strong performance on one family does not conceal a weakness on another.

## Keep manual review where judgment changes the outcome

Automation applies defined checks repeatedly and brings selected evidence into the review queue. Reviewers can reconcile that evidence with the application, request an explanation and resolve contradictions using independent records. Their work changes when software supplies a specific concern rather than an undifferentiated alert.

An illustrative lending workflow shows the division of responsibility. A statement produces a signal around its closing balance. The underwriter reviews the original file and the highlighted area, then obtains an authorized source record when the concern warrants it. A benign conversion may explain the signal; a mismatch with the source record may support escalation. Neither outcome was established by the detector alone.

![Document signals and source records feed lender policy, followed by an underwriter’s documented decision.](/blog/choosing-document-fraud-detection-software/lending-review-workflow-25d57b0d9d08.png)

Routing policy needs an owner and a defined action for incomplete evidence. A complete result without strong manipulation signals can continue through the existing controls when policy permits. Conflicting evidence goes to review or additional evidence collection. An unreadable submission needs a supported replacement or another review path, rather than automatic clearance.

The same boundary applies to insurance: a suspicious repair invoice can justify investigation without deciding whether a covered loss occurred. The adjuster still owns that decision.

## Integration includes the failure path

An application programming interface (API) is a connection method. Operational fit depends on what happens after the request arrives: how results attach to the right application, how duplicate requests are handled, what happens during an outage, and where incomplete analysis goes.

A deployment proposal needs documented behavior for retries, processing limits, peak intake and model-version changes. Batch review and an intake-time check serve different decisions. Re-screening a portfolio may support investigation, while an origination check must arrive before the relevant funding decision.

Documents also contain sensitive financial and identity information. Commercial and security review should establish processing locations, retention and deletion, access controls, subprocessors, training-use terms and the responsibilities of both parties. Hosted and on-premises options need evaluation against those requirements; a deployment label alone proves little about security.

Redaction needs particular care. It changes the submitted artifact and can remove evidence the analysis needs. An approved handling plan defines which version is analyzed and how it relates to the original, without collecting or retaining more sensitive data than the purpose requires.

## Compare total workflow cost

The commercial comparison needs to account for more than software charges:

**Total cost = software charges + integration and operating work + review and follow-up effort.**

A quote needs a clear billing unit, including whether pages, repeated submissions, retries or failed analysis incur charges. Minimum commitments, support and deployment fees also belong in the comparison. A low unit price can still create an expensive workflow if alerts require extensive follow-up.

Scam AI document detection is arranged with our team and priced by document and page. It is separate from the public self-serve media credit schedule. Document buyers therefore need a proposal matched to their intake, rather than an assumption that image credits cover document-forgery checks.

## Bring the decision into the demo

We also analyze submitted photos and video for signs of manipulation and AI generation. When an application also depends on a property or asset inspection, [Verifit guided remote inspections](https://scam.ai/verifit) provide a separate evidence-collection workflow. Customers follow a guided capture process, and the lending team reviews photos and verification results. Document analysis and remote inspection address different parts of the same decision; neither establishes income or replaces underwriting.

The purchase should leave your team with a defined scope, reviewable results and an accountable path for uncertainty. [Book a Scam AI demo](https://scam.ai/contact) focused on your document mix and the origination decision you need to support.

## Deepfake detection for AI agents: Scam AI × Aident
URL: https://scam.ai/resources/blog/verify-before-you-act-aident
Category: Partners
Published: 2026-10-02T12:00:00.000Z

[Scam AI × Aident](/blog/verify-before-you-act-aident/scamai-x-aident.svg)

*By Xingyu (Alex) Shen, Founding Engineer, Scam AI*

Scam AI and [Aident](https://aident.ai) show how an AI agent can check the photos, videos and recordings in a case before it relies on them. Aident Loadout lets the agent open the apps that hold the case, under the team's policies. Scam AI checks each supported media file for signs of AI generation or manipulation. The claims team makes the final decision.

An agent can summarize an insurance claim in minutes, but it takes every photo and video in the packet at face value. If one was generated or edited, the agent may rely on it in both its summary and its recommendation. Insurance fraud costs American consumers at least $308.6 billion a year, by the Coalition Against Insurance Fraud's count ([CAIF](https://insurancefraud.org/fraud-stats/)).

The workflow needs both halves. An agent that cannot reach the case has nothing to check, and one that reaches it without checking passes generated media straight into its brief. We call the joint workflow **Verify Before You Act**.

Below, a synthetic claim packet shows real detection results for a sample photo and video, run on our production API. The brief around those results shows how the agent routes them; it illustrates the workflow rather than recording a full Loadout run.

**At a glance**

- **Who:** Scam AI checks images, video and audio for AI generation and manipulation; [Aident Loadout](https://aident.ai) governs which apps and credentials an agent may use.
- **The use case:** before an agent summarizes a claim or recommends a next step, it checks the supported media in the packet, and a person reviews anything flagged.
- **The split:** Loadout governs how the agent reaches the case, Scam AI checks each file, and a person on the team decides.
- **The demo:** a synthetic claim packet; both media files were flagged on our production API, and the raw detection output is linked below.
- **Available today:** the Scam AI MCP server with its starter Skill, and Aident Loadout. We set up the joint workflow with each team.

## A sample claim, checked file by file

Sample case, synthetic files, not a real claim. The packet holds an AI-generated photo of the insured shop front, an eight-second clip made from one of our AI-generated faces, and a repair-estimate PDF.

[Evidence Authenticity Brief for a sample insurance claim: the shop-front photo and the video statement are both LIKELY\_AI with all eight video frames above the line, the repair-estimate PDF is unchecked, and the agent recommends holding the claim for the claims team](/blog/verify-before-you-act-aident/evidence-brief.webp)

*The verdicts are real results from Scam AI's public API on 2 October 2026 (UTC); the next steps show how the starter Skill routes them.*

Both media files came back `LIKELY_AI`, and all eight scored frames of the clip sat above the threshold. The API returned the same `verdict` and one-sentence `summary` that the MCP tool gives an agent.

Under the Skill, each flagged file gets a next step: ask for the original photo or send an inspector, and arrange a live video call with the claimant. The repair estimate stays unchecked, because this check covers images, video and audio only. Document forgery detection is a separate Scam AI plan, and our [TinyFish claims review](/resources/blog/reach-it-prove-it-tinyfish) shows that side.

Raw detection output and provenance: [photo result](/blog/verify-before-you-act-aident/demo/storefront-photo.production.json), [video result](/blog/verify-before-you-act-aident/demo/video-statement.production.json), [how the samples were made](/blog/verify-before-you-act-aident/demo/SOURCES.md).

## How Verify Before You Act runs

1. **The team opens the case** and gives the agent its review criteria.
2. **The agent gathers the files through Aident Loadout**, reading the case from the team's apps without seeing a raw credential.
3. **Scam AI checks each media file.** One `detect_media` call returns a verdict, a score and a one-sentence summary.
4. **The agent writes the brief, following a Skill**: one rule per verdict, one line per file, and the reason each flagged file needs a person.
5. **The team decides.** A person reads the brief, asks for what is missing and makes the call.

[Five stages from case packet to decision: the team opens the case, the agent gathers the files through Aident Loadout, Scam AI checks each file, the agent branches and writes the brief, and the team decides](/blog/verify-before-you-act-aident/verify-before-acting.svg)

*Cyan is Aident Loadout, teal is Scam AI's check, and white is the team's own steps.*

## How to use a deepfake-detection MCP server in an AI agent

MCP, the Model Context Protocol, is an open standard that lets an agent call real tools. Connecting Scam AI takes one config entry: the agent's MCP client points at `https://mcp.scam.ai/mcp` with an API key in the `x-api-key` header, or runs `npx -y @scam-ai/mcp-server` locally. Either way a check costs the same as a direct API call ([MCP & Skills](/platform/mcp)).

The agent then calls `detect_media` once per file, with a link or the file's bytes, or a local path when the server runs on the agent's machine. The tool accepts images, video and audio and works out which it received.

Our MCP page also publishes a starter Skill, `verify-media-before-acting`, a short instruction file with one rule per answer. It continues on `LIKELY_REAL`, lists `ALERT` for human review, escalates `LIKELY_AI` and treats a failed call as unchecked. Written into a Skill, the rule loads with every task instead of depending on how someone phrased the request.

## Where Aident Loadout fits

Loadout gives agents the integrations, credentials, policies and audit trail they need without handing them raw provider credentials ([Loadout overview](https://docs.aident.ai/loadout/overview.md)). API keys sit in an admin-managed Vault, and the audit shows each action with its status and the credits it used ([Vault and Audit](https://docs.aident.ai/loadout/vault-and-audit.md)).

In Verify Before You Act, Loadout is how the agent reaches the claim inbox or shared drive where the packet sits, under the team's policies. In this walkthrough, Loadout and Scam AI's MCP server are two connections in the same agent. Scam AI checks media for signs of AI generation or manipulation; Loadout governs what the agent may access.

## What does the agent get back for each file?

Every answer carries a `verdict` to route on, a `score` for the probability that the file is AI-generated or manipulated, and a one-sentence `summary` ([reading the result](/docs/reading-the-result)). The score is not a confidence level, and it is not proof ([MCP & Skills](/platform/mcp)). A video adds a per-frame series: about one frame a second up to 20 seconds, and 20 frames across a longer clip. Audio adds scored windows and answers only `LIKELY_REAL` or `LIKELY_AI`.

A failed check is also an answer: the server returns one of 13 stable error codes, so the agent branches on the code, never on the wording ([package reference](https://www.npmjs.com/package/@scam-ai/mcp-server)).

`unprocessable_media` is never charged, and the brief lists that file as unchecked. After a `timeout` the check may have run and been charged, so the agent reads `get_usage` before resending. **An unchecked file stays unchecked; it never becomes a clear result.**

## The same check in lending, onboarding and hiring

Claims are the first example, not the only one. A lending agent can check the selfie or ID video in a [loan application](/solutions/loan-credit) before it scores the file. An onboarding agent can route [KYC](/solutions/kyc-idv) selfies by the same three verdicts. A recruiting agent can check a recorded interview before it summarizes the candidate for a [hiring](/solutions/hiring-fraud) panel.

## Setting it up for a claims team

The Scam AI MCP server and its starter Skill are live on our [MCP page](/platform/mcp), and Aident Loadout is live. A team brings its review policy for the Skill's rules, the apps Loadout may open, and a place to hold the key. Our [insurance claims](/solutions/insurance-claims) page describes the review this flow feeds.

## FAQ

**How do I use a deepfake-detection MCP server in an AI agent?** Add the server to the agent's MCP configuration and call `detect_media` once per file. Then give the agent one rule per verdict in a Skill: continue on `LIKELY_REAL`, send `ALERT` to a person, and block or escalate `LIKELY_AI`.

**Does the agent decide whether a claim is fraudulent?** No: Scam AI reports a probability about each file, the brief sorts the results, and the claims team decides.

**Which files can the Scam AI MCP server check?** Images up to 10 MB, video up to 100 MB and audio up to 50 MB, or a link to a public video page or media file. Document forgery detection is a separate plan.

**What does a check cost?** One credit per image, per scored video frame up to 20 a clip, and per started minute of audio, at $0.02 a credit. New accounts start with 100 credits ([MCP & Skills](/platform/mcp)).

---

Agents that act on files customers send in should check them first. [Contact us to integrate the Scam AI MCP server and starter Skill into your agent workflow](/contact?from=blog-verify-before-you-act-aident).

Learn more about Aident: [Aident Loadout](https://aident.ai) · [Loadout documentation](https://docs.aident.ai/loadout/overview.md)

**Scam AI × Aident · 2026**

## How Scam AI and TinyFish vet candidates and claims
URL: https://scam.ai/resources/blog/reach-it-prove-it-tinyfish
Category: Partners
Published: 2026-09-22T12:00:00.000Z

[Scam AI × TinyFish](/blog/reach-it-prove-it-tinyfish/scamai-x-tinyfish.svg)

*By Xingyu (Alex) Shen, Founding Engineer, Scam AI*

Scam AI and [TinyFish](https://tinyfish.ai) have combined two checks into one review for hiring and insurance claims. TinyFish finds and reads the public web pages behind a candidate or a claim. Scam AI checks the submitted file, ties each page to the case and writes down what remains unverified. Reality, our review workspace, puts all of it in front of the person who decides.

Both decisions need both checks. Before a job offer goes out, a recruiter wants to know that the candidate's background holds up and that the interview video is genuine. Before an adjuster approves a payout, they want to know that the repair estimate came from the business named on it and that nobody changed the amount.

A web search can find a candidate's projects or a repair business's website; a file detector can find signs that a video or a PDF was altered. Neither answers the whole question alone. The two scenarios below are the ones we are filming; the recorded runs after them are the evidence so far.

**At a glance**

- **Who:** Scam AI detects altered and AI-generated video, images and documents; TinyFish builds web infrastructure for AI agents.
- **What:** a review workflow for two decisions: send the offer, pay the claim.
- **The split:** TinyFish retrieves and reads public pages. Scam AI checks the file and organizes the evidence. A named reviewer decides.
- **Proof:** two recorded runs with public receipts: seven pages bound and five résumé statements supported for my own résumé; three official sources answering for a sample vehicle claim.
- **The principle:** every finding shows its source, every gap stays visible, and the next check is named.

## Before the offer: the candidate and the interview

A hiring team is ready to make an offer to a remote engineering candidate. The recruiter has the résumé, the candidate's links, and an interview recording the candidate agreed to have checked.

TinyFish starts from those links. Its [web search API](https://tinyfish.ai/search) returns results, each with a URL, a title and a snippet. Its [web fetch API](https://tinyfish.ai/fetch) loads the pages our selection policy picks in a real browser and returns their text. The selection policy is a versioned rule for which kinds of source may answer which statement.

**TinyFish retrieves the page; whether it belongs to this candidate, and what it supports, is our question.** A namesake's page stays separate, and a page that cannot be found does not count against the candidate.

The interview recording gets its own check. The video file goes to the [Eva V1.6 Detection Model](/platform/eva), which reports whether the footage shows signs of a face swap or other manipulation. That result is about the file: it cannot say who the speaker is, just as the web results cannot say whether the video was altered.

The recruiter sees both sets of findings and the one open question: the candidate's identity. Next comes a live call with an independent identity check; the recruiter, not a score, decides whether the offer goes out. That is the [hiring fraud](/solutions/hiring-fraud) review we built this for.

## Before the payout: the repair business and the estimate

An adjuster has a repair estimate for **$40,180.00** and is about to approve it. In our sample, the original total was **$4,180.00**. We added one digit and left every line item in place.

TinyFish's part is the issuer. From the approved business details, Search and Fetch find its website or public records, which show that it exists and what it says about itself. They cannot prove that this PDF came from it.

Scam AI's part is the file. The document detector examines the exact bytes submitted, including the PDF's structure, metadata and revision history, for signs of editing after issue.

The adjuster sees the public context, the file result and the missing verification. The next action is to request the original from the issuer through an approved channel: a concrete step for an [insurance claims](/solutions/insurance-claims) reviewer before deciding on payment. Both sample PDFs are here: [original](/blog/reach-it-prove-it-tinyfish/receipts/samples/estimate-original.pdf) and [altered](/blog/reach-it-prove-it-tinyfish/receipts/samples/estimate-altered.pdf).

## What we have run so far

The runs below are different cases from the two scenarios; every number links to its receipt. Both exercise the web half of the review; a public file-check receipt is still to come.

### A public-context review of my own résumé

I authorized a review of my own résumé, starting from two anchors I supplied: my GitHub handle and my personal domain. Seven Search calls returned 46 results; our policy picked seven pages and Fetch read them all: my GitHub profile and a repository, my portfolio, an arXiv paper, Duke and DKU pages and my Google Scholar profile ([search receipt](/blog/reach-it-prove-it-tinyfish/receipts/hiring/public-context-receipt.json)).

**Each page earned its place through a route it can show, never through my name.** The GitHub profile links back to my domain and lists the same Scholar ID; the portfolio links the same projects; the Duke, DKU and arXiv pages carry the same project titles. MiniMax-M3 wrote that linkage down with citations, and Reality passed all seven sources on provenance and attachment ([packet](/blog/reach-it-prove-it-tinyfish/receipts/hiring/review-packet.json)). The same rule kept out directory pages listing more than 100 other people named "Xingyu Shen" and a name-only Scholar result.

The result: five résumé statements checked, five supported by cited sources, each with its limit stated: an author listing shows association, not the scope of a contribution ([report](/blog/reach-it-prove-it-tinyfish/receipts/hiring/report.html)). No image, video or PDF detector ran here. Its conclusion, as printed:

> Enough to choose an interview; not enough to choose an offer.

### Public vehicle context for a sample claim

A sample claim supplied a 2011 BMW X3's year, make, model and a partial VIN pattern from training material, never a complete VIN. [Three Search calls](/blog/reach-it-prove-it-tinyfish/receipts/vehicle/vehicle-context-receipt.json) returned 20 results and Fetch read three official sources. [NHTSA's vehicle decoder](https://vpic.nhtsa.dot.gov/api/vehicles/DecodeVinValuesExtended/5UXWX7C5%2ABA?format=json&modelyear=2011) identified the model as an X3 xDrive35i and flagged the incomplete VIN; [NHTSA's recall list](https://api.nhtsa.gov/recalls/recallsByVehicle?make=BMW&model=X3&modelYear=2011) for the model year returned four records; a BMW page gave the 300 hp specification.

**That is what the public web can say about a model, and no more.** The [readiness record](/blog/reach-it-prove-it-tinyfish/receipts/vehicle/vehicle-decision-readiness.json) binds it to the case with its limits. With a complete VIN, the vehicle's own records arrive through the carrier's first-party lane, where the VIN never touches a public route. The sample had neither. The record marked the claim ready for triage but not for payment or denial, and listed the six evidence groups still needed, the full VIN and the carrier's history first. The [report](/blog/reach-it-prove-it-tinyfish/receipts/vehicle/report.html) ended:

> NOT PAYMENT-READY. Continue intake.

The adjuster gets a precise request list rather than a guess.

Two held-out runs, a second résumé and a Kia claim, behaved the same way; those receipts are internal.

## What the reviewer receives

The output is a readable report and a structured record with the same findings; the [hiring packet](/blog/reach-it-prove-it-tinyfish/receipts/hiring/review-packet.json) above shows the format. A reviewer can answer four questions without redoing the run:

- **What supports this statement?** The source, its URL, when we read it and what it actually says.
- **Is it about the right person or case?** How we matched the source, with namesakes and conflicting information left visible.
- **What happened when the file was checked?** The exact file, the detector version, the result and its limits, or an explicit note that the check did not run.
- **What is still needed?** A live identity check, an issuer original, another document, or a look at conflicting evidence.

## How the Scam AI and TinyFish integration works

Six stages, one owner each:

1. **Open the packet.** *Customer.* The packet is the case as submitted: the authorized files, the permitted sources and a named reviewer.
2. **Compile claims.** *Parser and policy.* Each statement to check becomes its own row, with the supplied links and identifiers extracted. Nothing private, a file's contents or a full VIN, goes into a public query.
3. **Acquire.** *TinyFish.* Search finds candidate sources; our selection policy decides which ones Fetch reads.
4. **Verify.** *Reality.* Each source is bound to the case by its hash and tested for whether it belongs to this person or vehicle and what it supports.
5. **Check the file, organize.** *Reality and MiniMax.* Authorized files go to the detector. When the MiniMax-M3 language-model step runs, it organizes rows that already carry a citation and cannot add facts.
6. **Review.** *Customer.* The reviewer reads the evidence and its limits, then requests what is missing or decides.

[Six-stage review workflow, one owner per stage](/blog/reach-it-prove-it-tinyfish/six-stages.svg)

*Six stages, one owner each; orange marks TinyFish's step, teal a Reality check, white a human step.*

## Why does a run stop instead of scoring?

The résumé run shows how a page earns its place: through an anchor the candidate supplied or a link from one, never through a name alone. The vehicle run shows how a gap is handled: model-level facts are labelled, and the missing VIN becomes a named request. When a source is unavailable, an identity is ambiguous or a detector cannot run, the report keeps that gap visible.

The [system contract](/blog/reach-it-prove-it-tinyfish/receipts/system-contract.json) has one success state, `EVIDENCE_PACKET_READY`. It means the material is ready for human review, not that the candidate passed or the claim should be paid. It also names 15 reasons a run can stop:

`AUTHORIZATION_REQUIRED` · `INTAKE_INCOMPLETE` · `POLICY_OR_JURISDICTION_REQUIRED` · `CLAIM_TOO_BROAD` · `AMBIGUOUS_ENTITY` · `NO_TEXT_BRIDGE` · `SOURCE_UNAVAILABLE` · `SOURCE_VERSION_CHANGED` · `RIGHTS_BLOCKED` · `INSUFFICIENT_PUBLIC_EVIDENCE` · `CONFLICTING_EVIDENCE` · `MEDIA_UNRESOLVED` · `MODEL_SCHEMA_REJECTED` · `SELECTION_AUDIT_REJECTED` · `FIRST_PARTY_VERIFICATION_REQUIRED`

**A named stop makes an unfinished check explicit and never turns missing evidence into a finding about the person or the claim.** The contract also forbids outputting a hire, a rejection, a ranking or a payment decision.

## For engineers: how the evidence stays checkable

We call the saved record of each search or file check a *receipt*: what we sent, what came back and the status. A SHA-256 hash, a fingerprint of the saved bytes, pins down which exact response or file a result refers to. **It does not make the contents true.**

The first query in my hiring run was `"shenxingy" "alexshen.dev" "Scam AI"`: the handle and domain I supplied, plus my employer. The saved Search response keeps its positions and hash; each page selected for Fetch records why the policy chose it and whether retrieval succeeded.

Every source row carries nine required fields: `source_id`, `url`, `source_class`, `capture_time`, `selection_audit_id`, `selection_route`, `selection_reason`, `fetch_status` and `text_sha256`. A failed retrieval is still a row, with its status. The reviewer's decision sits in its own record and cannot change the evidence or its hashes.

## What's next for the Scam AI and TinyFish integration

Two short films will show the hiring and claims scenarios end to end; their detector and search captures are still to come. Sources behind a login, such as a carrier portal, can enter the packet through a separately authorized lane. That lane is TinyFish's [web agent](https://tinyfish.ai/agent) for [multi-step web tasks](https://tinyfish.ai/agent) or its [browser sessions for agents](https://tinyfish.ai/browser), under the same provenance rules; neither recorded run needed it.

If you review candidates or claims, [talk to us about running this on your own material](/contact).

*Read the other side of this story: [How scam.ai verifies hiring candidates and insurance claims with TinyFish](https://www.tinyfish.ai/blog/scamai-tinyfish-hiring-insurance-claims-verification).*

## FAQ

**Does the Scam AI and TinyFish integration decide whom to hire or which claim to pay?** No. It assembles evidence and names the open questions; the authorized reviewer decides.

**Does finding a profile verify a candidate's identity?** No. A matching name never connects a page to a person, and public context cannot replace an independent identity check.

**What happens when there is no public information?** The report records the gap and names the next step, a direct question or a first-party check; a limited web presence is not a negative finding.

**What happens to private data?** Full VINs, identity-document numbers and file contents stay out of public queries; the [system contract](/blog/reach-it-prove-it-tinyfish/receipts/system-contract.json) bars a complete VIN, owner identifiers, a policyholder's name and credentials from any public packet.

---

Learn more about TinyFish: [web search API](https://tinyfish.ai/search) · [web fetch API](https://tinyfish.ai/fetch) · [multi-step web tasks](https://tinyfish.ai/agent) · [browser sessions for agents](https://tinyfish.ai/browser)

**Scam AI × TinyFish · 2026**