Home/Guides/Document Fraud Detection
Document Fraud
AI document fraud detection that catches edited balances, AI-generated IDs, and everything in between.
AI document fraud detection means analysing a PDF, image, or scan for evidence of editing, forgery, or AI generation. TamperCheck runs structural forensics, pixel-level analysis, and metadata inspection on every document, then returns an authentic, suspicious or tampered verdict with the specific fields that changed and a calibrated risk score.
- No document storage
- Digital, scanned or photographed
- ~1 minute per verdict
Last updated:
Market trends
244%
year-on-year rise in digital document forgeries in 2024, which now outnumber physical counterfeits
Entrust 2025 Identity Fraud Report, Nov 2024444,000
fraud cases filed to the UK National Fraud Database in 2025, a record; over half were identity fraud
Cifas Fraudscape 2026, Mar 202689%
of companies in the Netherlands and Belgium were confronted with fraud, up from 78% a year earlier
Allianz Trade Fraud Trend Report 2026, Apr 2026£1.16bn
of fraudulent general insurance claims detected by UK insurers in 2024, across 98,400 claims
Association of British Insurers, Nov 2025The problem
Why document fraud detection has to be automated now
Document fraud has changed shape. Crude edits are still around, but the threat that matters now is documents that were never real at all. In Entrust's 2025 Identity Fraud Report, digital forgeries made up 57% of all document fraud in 2024, overtaking physical counterfeits for the first time after a 244% year-on-year rise.
The tools are cheap and public. Prosecutors say one service, OnlyFake, sold more than 10,000 AI-made fake IDs at around $15 each before its creator pleaded guilty in 2026. Expense platforms report the same shift: AppZen found AI-generated receipts went from none to roughly 14% of fraudulent receipts within a year.
Human review cannot keep pace. A reviewer checks whether a document looks right, and most modern fakes are built to look right. Automated document fraud detection checks what people can't see - PDF font tables, pixel-level noise, compression artifacts, arithmetic, metadata and AI-generation signatures - on every document, not a sample. It doesn't replace your compliance team; it gives them a forensic layer underneath.
In the news
Document fraud in the news
Feb 2026
OnlyFake creator pleads guilty in New York
Prosecutors say the AI fake-ID site produced at least 10,000 digital fake IDs, including passports for 56 countries, and took about $1.2 million in cryptocurrency.
US Attorney's Office, SDNYOct 2025
AI-generated receipts reach expense claims at scale
AppZen found AI fakes made up about 14% of fraudulent receipts in September 2025, up from none a year earlier. Ramp flagged more than $1 million in fake invoices in three months.
Computerworld, reporting the Financial TimesAug 2025
FBI and Dutch police seize the VerifTools fake-ID marketplace
The site sold fake identity documents marketed for getting past verification checks. The FBI linked about $6.4 million in proceeds to it.
BleepingComputerNov 2024
FinCEN warns banks about AI-generated identity documents
Its alert describes criminals using generative AI to produce fake driver's licences and passports to open accounts, and asks institutions to flag suspected cases in suspicious activity reports.
FinCEN
Types of document fraud
The four classes of document fraud
Each class fails in a different place, which is why no single check catches them all.
Edited documents
A genuine document with one or more values changed: a balance, a salary, a date. Most of the page is authentic, so it passes a visual check. Caught by pixel forensics, font and text-layer analysis, and arithmetic that no longer reconciles.
Spliced documents
Parts of different genuine documents combined: a real letterhead on a fabricated body, a seal or signature lifted from another file, a photo swapped on an ID. Caught by noise and compression mismatches between regions, and by duplicate detection across submissions.
Template-forged documents
Built from scratch on a template that imitates a bank, employer or registry. Usually internally consistent, because the template does the maths. Caught by PDF structure and producer metadata, layout deviations from the real issuer, and cross-field logic.
AI-generated documents
Images or PDFs produced by generative models or fake-document services, often staged as phone photos. Nothing was edited, because nothing was ever real. Caught by AI-generation signatures, capture provenance, and data-layer checks such as MRZ and barcodes.
AI-generated documents
How to detect AI-generated documents
An AI-generated document has no original - nothing was edited, so tamper checks that look for editing traces find nothing. Detection has to separate machine-generated output from institution-issued output, using six kinds of evidence.
Spectral noise
Image generators leave noise patterns that real cameras, scanners and printers don't. Every image region is read for them, including documents staged as phone photos.
Rendering fingerprints
Banking software, payroll systems and government printers each render text and graphics a specific way. Generated documents take different rendering paths that show up forensically.
Typography
Generated text has subtle kerning errors, uneven baselines and character shapes that match no real font. Character-level font analysis catches them.
PDF structure
Generated PDFs have object structures, font embedding and producer metadata that don't match how the claimed issuer produces its documents.
Data plausibility
Generated financial documents often contain arithmetic that doesn't reconcile, impossible date sequences, or identifiers in the wrong format for the issuer.
Compression and encoding
Generative tools encode and compress content differently from scanners and institutional PDF generators, and the difference persists after re-saving.
What we detect
Forensic signals that expose fraud
Structural PDF analysis
Object streams, font tables, embedded metadata, cross-reference tables - the internal structure of every PDF, inspected for the fingerprints editing tools leave behind.
Pixel-level forensics
Error Level Analysis, noise residue, JPEG ghost detection, and compression artifact maps reveal regions of an image or scan that have been edited.
AI-generation detection
Synthetic documents from generative AI carry spectral signatures and noise patterns that real scanners and printers don't produce. The CV layer is trained specifically on these.
Arithmetic & logical consistency
Running balances, subtotals, GST calculations, date sequences, and cross-field relationships are all verified. A single edited number breaks the chain.
Producer metadata fingerprinting
Genuine documents have consistent producer metadata (which software created them). Tampered files often show traces of editing tools that the original issuer would never use.
Visual layer vs text layer divergence
Edit a PDF and the rendered pixels can drift from the underlying text layer. The mismatch is invisible on screen and obvious at the character level.
Manual vs automated
Manual review vs automated document fraud detection
Automation doesn't take people out of the decision. It decides which documents need them.
| Manual review | Automated detection | |
|---|---|---|
| What it checks | What the document says and how it looks | How every part of the file was made, plus what it says |
| A single edited field | Missed unless the edit is visibly clumsy | Flagged from pixel, font and text-layer evidence |
| AI-generated documents | Built to pass a visual check | Detected from generation and provenance signals |
| Coverage | A sample, when the queue allows | Every document, at submission |
| Time per document | Minutes, and it varies by reviewer | Seconds, and it is consistent |
| Audit trail | Reviewer notes | Named findings, a risk score and a shareable report |
Buying guide
How to evaluate a document fraud detection vendor
Demos run on the vendor's best samples. Test these six things on your own documents instead.
Findings, not just a score
Can the tool show which field changed and where, in words a reviewer can act on? A bare risk number is hard to defend in a dispute or an audit.
Coverage of your documents
Test the document types and countries you actually receive, including phone photos and scans, not only clean PDFs.
AI-generated detection
Ask how it handles documents that were never edited because they were generated. Template matching alone will not catch them.
Speed and integration
Check latency at your volume, webhook support, and whether results can be fetched by your own document ID.
Data handling
Find out whether documents are stored, for how long, and whether they are used to train models.
Pricing that fits your volume
Per-document pricing without minimums makes a pilot cheap. Check how the price scales at production volume.
Use cases
Where document fraud detection is used
Lending and mortgages
Bank statements, pay stubs and tax returns checked for inflated income before underwriting.
KYC and onboarding
IDs, passports and proof of address checked for edits and AI generation before an account opens.
Insurance claims
Invoices, receipts, repair estimates and medical bills checked before payout.
Tenant screening
Pay stubs, bank statements and references checked before a lease is signed.
HR and background checks
Degrees, certificates and employment letters checked before an offer is made.
Expenses and accounts payable
Receipts and supplier invoices checked for edits, AI generation and reuse before money leaves.
What you get back
An example API response
One upload returns a verdict, a calibrated risk score, named findings and a plain-language summary. Results arrive by webhook, or can be fetched by job ID or by your own document identifier.
GET /api/v1/documents/{id}/
{
"id": "550e8400-e29b-41d4-a716-446655440000",
"status": "completed",
"strategy": "tampering_detection",
"document_identifier": "loan-app-12345",
"document_type": "bank_statement",
"verdict": "tampered",
"risk_score": 42,
"confidence": 0.78,
"page_count": 1,
"findings": [
{
"category": "font_inconsistency",
"severity": "medium",
"description": "Mixed font metrics in the totals row.",
"confidence": 0.71
}
],
"human_summary": "The closing balance is set in different font metrics from the rest of the statement, consistent with an edit after export.",
"completed_at": "2026-04-02T12:00:03Z"
}The solution
One endpoint, one upload, one verdict per document
TamperCheck is an API-first document fraud detection tool built for B2B teams that process documents at volume - lenders, insurers, KYC providers, HR teams, and claims adjusters. One endpoint, one upload, one verdict per minute.
How it works
- API-first - integrate in under 30 minutes; webhooks for async results
- 200+ forensic checks per document, automatically
- Findings your team can act on - not just a risk score
- Risk score (0-100) calibrated per document type
- per-document pricing - no subscriptions, no minimums
- Zero document storage - files processed ephemerally
- Works on any document type - bank statements, payslips, IDs, invoices, certificates
Why us
What you won't get elsewhere
Findings, not a score
We name the field that changed and where it sits on the page.
Others hand you a number to justify.
One endpoint, every type
100+ document types, 190+ countries, no per-type setup.
Elsewhere, each new form is a project.
Catches what was never real
A dedicated pass for AI-generated and deepfake documents.
Template matching only catches edits.
Nothing is retained
Analysed in memory, then discarded. Never trained on.
Most services keep documents for QA.
Live in minutes
Published pricing, no minimum. Test it before you talk to us.
Platforms want a contract first.
Sits alongside your stack
Keep Onfido or Jumio for face match. We add the forensics.
Not a rip-and-replace.
FAQ
Common questions
What is document fraud detection?
Analysing a document file for evidence that it was edited, forged, or generated by AI. The techniques - PDF structural analysis, pixel-level error analysis, font metric checks, metadata inspection, AI-generation detection - catch edits that aren't visible to the human eye.
What is AI document fraud detection?
Software that uses machine learning, computer vision and forensic analysis to decide whether a document is genuine - catching edited fields, forged templates and fully AI-generated documents. TamperCheck combines 200+ forensic checks with an AI adjudication layer and returns a verdict with the specific findings behind it.
Which AI tools can be detected?
Detection targets the forensic properties of the output, not a specific tool. It works on documents from general image models, chat assistants and dedicated fake-document services alike - anything that leaves machine-generation signatures in the file - including models released after the detector was built.
Is document detection the same as document fraud detection?
Different problems. 'Document detection' usually means finding and reading a document in an image - that's OCR and document capture. Fraud detection asks whether the document is genuine. Most workflows need both: OCR to extract data, fraud detection to decide whether that data can be trusted.
What are the main types of document fraud?
Four kinds: edited documents (genuine document, changed values), spliced documents (parts of different documents combined), template-forged documents (built from scratch to imitate an issuer), and AI-generated documents (produced by generative models, never real). Each fails in a different place, so detection needs several layers.
What's the best document fraud detection tool?
Depends on your volume and integration model. TamperCheck.ai is built for B2B teams that need API-first fraud detection across bank statements, payslips, IDs, and invoices, no subscription, zero document storage, verdict with named findings.
How accurate is automated document fraud detection?
Ask any vendor for results on your own documents rather than a headline benchmark - accuracy depends on document type, fraud kind, and file quality. TamperCheck shows the evidence behind each verdict: the field, the signal, the confidence. Your team checks the reasoning and sets the risk threshold.
Can document fraud detection catch AI-generated forgeries?
Documents from generative AI tools carry spectral signatures, noise patterns, and compression artifacts that differ from scanner- or printer-produced documents. TamperCheck's computer vision layer targets these signatures, including outputs from current-generation diffusion and transformer models.
How does document fraud detection integrate with our KYC workflow?
Single REST endpoint. Submit a document, get a verdict, branch your workflow on the result. Most customers slot it in after OCR and data extraction but before human review - suspicious documents get flagged so reviewers focus on the high-risk cases.
What does it cost to try document fraud detection?
Add your first top-up and TamperCheck matches it in free credits, so you start with double the balance. Every check covers the full analysis with a complete report. No subscription, and unused credit stays in your wallet.
Or skip the call entirely
We'll match your first top-up in credits, so you can run your own documents through it now.