Inscribe says AI is under 5% of document fraud, AppZen 70.8%. Both are right.

Two companies that detect document fraud published numbers this year that
appear to contradict each other completely.

Inscribe (20 January 2026, 2026 State of Document Fraud Report):
AI-generated documents were less than 5% of the fraudulent documents it
flagged in lending during 2025. Their words: "the majority of document fraud
is still template-based."

AppZen (26 June 2026, via PYMNTS):
70.8% of flagged fraudulent expense receipts were AI-generated by mid-May
2026: up from 0% in March 2025. That's 1,471 AI-generated fake receipts,
submitted by 745 employees, at 174 companies.

Less than 5%. Or 70.8%. Pick your headline.

Both are right

The numbers measure different populations, and once you see the denominators
the contradiction disappears.

Inscribe's denominator is every fraudulent document flagged across its
lending customers: bank statements, pay stubs, tax forms, business filings.
That's a world where forgery already worked before generative models arrived.
Inscribe's own measure of it: template-based fraud was 1 in 5 flagged
documents in 2025, up from 1 in 14 in 2024. The usual explanation, which I
went looking for a primary source on and could not find, is that template
farms sell editable pay stubs and bank statements that look right because
they started from the real thing; the share is measured, the explanation is a
hypothesis. Either way, against that installed base, AI generation is a small
new entrant. If you're underwriting loans, most of the fake documents you'll
see this year are still edited or templated, not generated.

AppZen's denominator is flagged fraudulent receipts in expense reports.
A receipt is the easiest document in the world to generate: low resolution
expected, no issuer database to cross-check, thousands of legitimate layout
variations. The base rate of sophisticated receipt forgery was always near
zero, because nobody needed sophistication to fake a $40 dinner. When image
models made generation free, it took over that category almost instantly, 0%
to 70.8% in fourteen months.

Same industry. Different documents, different fraud economics, different
numbers. Anyone quoting either figure without its denominator is misleading
you, and both figures are being recirculated stripped of exactly that
context. "71% of expense fraud is AI-generated" is already travelling through
AI-written content farms; it isn't what AppZen measured (70.8% of flagged
fraudulent receipts, not of expense fraud, not of receipts).

The number that should actually worry you

Here's the finding that doesn't depend on any vendor's platform data, from
AIForge-Doc (arXiv 2602.20569), an
academic benchmark of tampering detectors against AI-inpainted documents:

• DocTamper, a specialised detector, scores 0.563 AUC out of distribution, against 0.98 in-distribution. Trained defenses collapse on fakes they weren't trained on.
• TruFor: 0.751, against 0.96 in-distribution.
• GPT-4o as a detector: 0.509, a coin flip.

The detectors built for the previous era score at chance on the new one.
That is independently measured, it is not marketing, and it's the reason we
treat every vendor detection-rate claim (including any we might be tempted
to publish ourselves) as unverifiable until someone shows a methodology.

What this means if you handle documents

1. Ask for denominators. "X% of fraud is AI" means nothing until you know fraud of what, flagged by whom, out of what population.
2. Match the defense to the document. Expense receipts and bank statements are different threat models. One is being flooded by generation; the other is still dominated by editing and templates, which leave structural traces (revisions appended after creation, producer strings that changed, metadata that disagrees with itself) that don't depend on how the visual layer was made.
3. Distrust round detection rates. Nobody in this market publishes a false-positive rate on legitimate documents. We publish ours, measured on uncurated public document populations, engine-version-stamped: tamperlens.com/evidence. We'd genuinely like not to be the only ones.

Tamperlens reads document structure and returns fraud signals as evidence, never an approve/deny verdict. Free checker, no signup:
tamperlens.com.