Reading AI detection results

AI detection results combine a verdict, a score and, for video or audio, evidence from the analyzed moments. These fields describe what Eva V1.6 detected in one run. They do not establish authenticity, manipulation, intent or wrongdoing on their own.

Treat the response as a probabilistic signal within your own policy. A Scam AI result must never be the sole basis for a consequential decision about a person. Use independent evidence and an accountable review process for those decisions.

Verdicts are routing signals#

verdict is a compact interpretation of the model output. It is not a pass, block or final case decision.

The shared response enum includes all three verdicts. The current audio path returns LIKELY_REAL or LIKELY_AI. Clients should still handle ALERT safely if it appears in an audio response, usually by sending it to review.

The score is a likelihood signal#

score is a number from 0 to 1 representing the model's estimated likelihood that the media is AI-generated or manipulated. Values closer to 1 carry stronger AI or manipulation signals. Values closer to 0 carry weaker signals.

The score does not represent:

• the API's accuracy on this request;
• certainty that the verdict is correct;
• the percentage of the file that was generated or altered; or
• evidence about who created the media or why.

Use score to order or prioritize cases only within a policy your organization has defined for that workflow. Do not copy a threshold from another use case. The cost of an error, the available supporting evidence and the action's reversibility all affect where review belongs.

For video, threshold_used is the reference line applied to the per-frame series for that run. It describes model behavior. It is not your organization's approval, decline or enforcement threshold.

Evidence depends on the media type#

The headline verdict and score summarize the run. Evidence fields show which sampled moments contributed information and which could not be scored.

Null, absent and abstained have different meanings#

Never convert a null value to 0. Zero means the model produced a low manipulation score. Null means no score was produced.

• score: null means no headline likelihood is available. Route the result through your unscored or review path rather than treating it as LIKELY_REAL.
• frames[].score: null means that video frame was not scored. Read frames[].cause, such as timeout or not_attempted, and keep the gap visible.
• segments[].value: null with status: "abstained" means the audio window was not scored. reason_codes explains why, for example no_voice_content.
• An absent modality field can be normal. Images do not return frames[] or segments[]; uploads do not return source; and a video can omit threshold_used or frames[] when there is no series to show.
• Nullable media metadata such as filename, mime_type or bytes means that value was not known. It does not change the detection score.

When no detector can assess the media, the API can return 422 NOT_ASSESSABLE instead of a completed detection. The error's reasons[] describes the abstention. Audio reasons can include no_voice_content or audio_too_short. Treat this as an explicit lack of evidence, not as a real or AI result. The error reference documents the response shape.

Keep the detection separate from your decision#

An illustrative policy-led integration preserves the model output and records the customer's action as a separate event:

detection signal = verdict + score + evidence + evidence gaps
policy input      = detection signal + case context + independent evidence
policy outcome    = review, step-up, label, hold or another defined action

Before applying a policy, confirm that the response refers to the expected media, review any null or abstained evidence, and check link coverage. Store id, verdict, score and relevant evidence with the policy version and the resulting human or system decision. Do not rewrite the detection verdict to match the later outcome.

Supporting response fields#