Industry whitepaper / October 2026
The Synthetic
Evidence Report
Lending Edition
How AI-fabricated documents are rewriting loan-origination fraud. And what lending teams can do about it.
43 pages / 10 chapters / 43 sources
/ Public-evidence research
From the public record
1 in 119
U.S. mortgage applications with fraud indications
Q2 2026 · Model-scored indications, not confirmed fraud
Cotality · Report [1]46%
of Fannie Mae’s investigative findings involved income misrepresentation
2025 loans · Share of investigative findings
Cotality, citing Fannie Mae · [1]+267%
rise in first-party auto-lending fraud losses
Q3 2018–Q3 2025 · TransUnion’s credit data
TransUnion · Report [2]These datasets measure different things. None establishes the share of document fraud caused by AI.
The report in brief
The data validates.
The document is fabricated.
A pay stub can add up. A bank statement can reconcile. An employer can answer the phone. The file behind every one of those checks can still be fabricated.
This report examines synthetic evidence across mortgage, auto, consumer, and small-business lending. It connects the public record to the work of intake, underwriting, risk, and quality-control teams.
“The document is the attack surface.”
The Synthetic Evidence Report · Lending Edition
Inside the report
- 01How fabricated artifacts exploit the gap between identity checks and income validation.
- 02What regulators, credit bureaus, and court records actually establish.
- 03A five-play defense framework, a vendor checklist, and a diagnostic for your own files.
A field guide, cover to cover.
43 pp.Unlock once to read every chapter and download the PDF.
- 01Executive summaryp. 03
- 02The anatomy of synthetic evidencep. 06
- 03The evidence: this is already happeningp. 10
- 04Why incumbent origination controls failp. 15
- 05The stakes: the liability outlives the loanp. 18
- 06The defense framework: five playsp. 22
- 07The buyer’s checklistp. 26
- 08Case studies: the public recordp. 29
- 09How Scam.ai detects synthetic evidencep. 34
- 10The diagnostic: test us on your own filesp. 36
Plus four appendices: evidence typology, regulatory timeline, methodology, and the complete source list.
A closer look
Readable isn’t the same as real.
Explore three documents a lender sees every day. Each illustrates why extracting correct data leaves an authenticity question unanswered.
EARNINGS STATEMENT
- Document ID
- SAMPLE-026
- Net pay
- $4,280.00
✓ The math adds up.
Illustrative sample. These are invented values, not a scan result.
The authenticity gap
A generator can calculate payroll deductions correctly. That doesn’t prove an employer issued the file.
Examine template reuse, fonts, layout, and PDF structure.
Report §2.3 · Fannie Mae [15], Freddie Mac [16]
Evidence, with its limits attached
Built on the public record.
Research compiled in October 2026 from regulators, law enforcement, courts, GSEs, credit bureaus, data firms, and journalism.
Every number has an evidence class. Detected signals, confirmed outcomes, surveys, estimates, and regulatory actions remain distinct.
The report excludes detection-vendor research. Its cited data and quality-control firms also sell analytics; their figures describe their own datasets. The full methodology and all 43 references are included.
What the public evidence cannot tell us
No regulator, GSE, or neutral data firm publishes the share of income-document fraud that is AI-generated rather than hand-edited. Prosecutions establish attack patterns; they do not, by themselves, establish AI prevalence.