A deepfake document looks completely real. It just never existed.
A deepfake video can fool the eye for a few seconds. A deepfake document can fool a lender, landlord, or admissions office for a lot longer. As generative AI models improve, a growing share of fraudulent submissions aren't edited originals at all - they're synthetic from the first pixel. TamperCheck's job is to tell the difference.
The problem
Deepfake documents are a different problem than tampering
Most document fraud starts with a real document that gets edited - a balance changed, a name swapped, a date altered. A deepfake document is different: there is no real original. The entire bank statement, ID, or certificate was generated from a text prompt or reference image by an AI model, complete with plausible branding, formatting, and data that was never issued by any real institution.
These are produced with image-generation models, AI form-fillers, and template-editing tools that can assemble a convincing document in seconds - no design skill or Photoshop experience required. The barrier to producing a passable fake has effectively dropped to zero.
Standard onboarding and review workflows aren't built to catch this. Liveness checks confirm a live person is present, not that their submitted document is genuine. OCR reads what a document says, not whether it was ever real. Both can pass a deepfake document without raising a flag, because neither one is actually checking for synthetic origin.
What we detect
Forensic signals that expose fraud
No genuine original to compare against
Tamper detection often works by finding inconsistencies against an expected baseline. Deepfake documents have no baseline at all - TamperCheck instead analyses whether the document's production signatures match any real-world issuing process.
Spectral and noise signatures
AI image generators leave characteristic noise patterns and spectral artifacts that real scans, photos, and institutional print runs don't produce. These are invisible to the eye but detectable forensically.
Template and issuer-format drift
Genuine institutions produce documents from a small set of fixed templates. AI-generated documents drift slightly from any real template - in spacing, alignment, or field structure - even when they look convincing at a glance.
Metadata that doesn't match a real pipeline
Real documents carry metadata fingerprints from the software, scanner, or printer that produced them. AI-generated files often carry no metadata, generic metadata, or metadata inconsistent with the claimed issuer.
Suspiciously clean (or subtly broken) internals
Generated documents sometimes get arithmetic, dates, or cross-field references slightly wrong in ways a real institution's system never would - or, conversely, look unnaturally tidy compared to real-world data.
Cross-document inconsistency
When a deepfake ID and a deepfake supporting document are submitted together, small inconsistencies between them - fonts, layout conventions, issuing-authority details - become a strong signal.
The solution
How TamperCheck tells a deepfake document from a real one
TamperCheck runs every upload through forensic checks built for both failure modes: edited originals and fully synthetic documents. The same API call covers both, so your team doesn't need to guess which kind of fraud it's looking at.
How it works
- Dedicated AI-generation and deepfake detection layer, run alongside traditional tamper checks on every document
- Works on PDFs, images, and scans - whatever format the deepfake was output in
- 130+ forensic checks spanning spectral analysis, typography, metadata, and structural fingerprinting
- Plain-English verdict in about a minute - AUTHENTIC or TAMPERED, with the specific signal that triggered it
- Model-agnostic - detection targets the forensic properties of the output, not a specific generator, so it holds up as new tools appear
- Zero document storage - documents are analysed in memory and discarded immediately after
FAQ
Common questions
See it working on your documents
Start with $5 in free credits - no contract, no card required. Upload your first document and get a verdict in about a minute.