# What Is a Deepfake? (And Why Deepfake Documents Are the Fraud You Haven't Prepared For) > What is a deepfake, and what's a deepfake document specifically? How AI-generated IDs, bank statements, and payslips are made, why KYC and liveness checks miss them, and what actually catches one. *Published 2026-06-29 · 9 min read · TamperCheck.ai* Canonical: https://tampercheck.ai/blog/what-is-a-deepfake-document --- A deepfake video convinces the eye for a few seconds before someone gets suspicious. A deepfake bank statement convinces a lender for a lot longer than that - long enough to fund a loan that should never have been approved. Both are "deepfakes." Almost everything written about them is about the first kind. This guide is about the second. ## What is a deepfake? A deepfake is synthetic media - an image, video, audio clip, or document - generated or manipulated by AI to appear authentic when it isn't. The term originally described face-swapped videos, and that's still the most visible example: a politician's face mapped onto someone else's body, a celebrity "saying" something they never said. What's changed is the range of things AI can convincingly fake. The same generative models that can produce a realistic face can now produce a realistic passport, a realistic payslip, or a realistic bank statement - complete with correct fonts, plausible logos, and numbers that add up. That's a deepfake document. ## What is a deepfake document? A deepfake document is a document - an ID, a bank statement, a payslip, a degree certificate - that was generated by AI rather than issued by a real institution and then altered. There's no real original behind it. The entire thing, branding and all, is synthetic. That's the distinction that matters most: a deepfake document is not the same problem as a tampered one. - **A tampered document** started life as something real. A genuine bank statement where someone has edited the closing balance. A real payslip with the net-pay figure changed. Most of the document is authentic; a few fields aren't. - **A deepfake document** never existed in any genuine form. An AI model was given a prompt - "a UK bank statement for a Barclays current account" - and produced a complete, self-consistent fake from nothing. Both can pass a quick visual check. Only one of them has a real document somewhere to compare against, which is why they need different detection strategies - see [Document Tampering and Fraud: Everything You Need to Know](https://tampercheck.ai/blog/document-tampering-fraud-complete-guide) for how the two categories map onto TamperCheck's forensic checks. - **71%** — of EMEA fraud professionals worried about deepfake threats. GBG Identity Fraud Report 2025 - **$81–98k** — median loss per synthetic identity fraud case. Federal Reserve Bank of Boston - **130+** — forensic checks TamperCheck runs per document, including dedicated AI-generation detection ## How are deepfake documents made? The short version: the same way any other AI-generated image is made, pointed at a document instead of a face. Image-generation models can be prompted directly for a document type and country, producing a complete, photorealistic image of an ID or financial statement. AI-powered template editors and form-fillers go a step further, auto-populating a convincing layout with plausible names, dates, and figures pulled from the prompt or a stolen identity. None of this requires design skill, access to specialist forgery tools, or even much effort - a usable fake can be produced in well under a minute. That's the part that's changed the threat model. Document fraud used to require some combination of skill, time, and editing software. Now it requires a prompt. ## Why standard KYC workflows miss them Most digital onboarding and verification stacks run two checks: liveness detection and document validation. Neither one is built to catch a deepfake document on its own. **Liveness detection** confirms that a real, physically present person is submitting the application - not a photo, video, or pre-recorded clip. It says nothing about whether the document that person is holding up or uploading is genuine. A real, live, willing applicant can still submit a synthetic bank statement. **Document validation**, in most stacks, means OCR plus template matching: does the layout look like a known format, and can the text be extracted cleanly? OCR reads what a document says. It doesn't check whether the document was ever issued by the institution it claims to be from. A deepfake document that matches a known template and contains internally consistent data sails through both checks. This is the same gap covered in more depth in [Liveness Detection vs Document Forensics](https://tampercheck.ai/blog/liveness-detection-vs-document-forensics) - the two checks solve different problems, and a complete verification stack needs both, not just one. > **WARNING:** Passing liveness and passing template matching tells you two things: a real person is present, and the document looks roughly right. Neither one tells you the document was ever genuine. That's a forensics question, not a liveness or OCR question. ## What actually catches a deepfake document Because there's no genuine original to compare a deepfake against, detection has to work differently than standard tamper-checking. It focuses on the production signatures a document carries - the fingerprints left behind by how it was actually made, not by what was changed on it. The forensic signals that matter most: - **Spectral and noise signatures** - AI image generators leave statistical noise patterns that differ from real camera sensors, scanners, and institutional print runs, even when the image looks clean to the eye. - **Template and issuer-format drift** - real institutions issue documents from a small, fixed set of templates. Generated documents drift slightly from any real template in spacing, alignment, or field structure. - **Metadata that doesn't match a real pipeline** - genuine files carry metadata fingerprints from the software or hardware that produced them. AI output often carries none, or metadata inconsistent with the claimed issuer. - **Cross-field and cross-document plausibility** - dates, arithmetic, and details that don't quite hold together internally, or that don't match a companion document submitted in the same application. Run together, these signals catch the majority of deepfake documents in about a minute per upload - faster than a human reviewer could even open the file. [TamperCheck](/) runs all of them as part of its standard 130+ check pass, alongside the checks that catch conventionally tampered documents, so one API call covers both failure modes. The full breakdown of the AI-generation side of that detection lives on [What Is a Deepfake Document?](https://tampercheck.ai/problems/what-is-a-deepfake-document) and, for the identity-document-specific version of this problem, [Detect Deepfake IDs](https://tampercheck.ai/problems/detect-deepfake-id). If the document in question is an image rather than a financial PDF - a photo ID, a selfie, a scanned certificate - the same underlying AI-generation signatures apply, with a few image-specific tells. [The 7 Things That Give Away an AI-Generated Image in 2026](https://tampercheck.ai/blog/deepfake-image-detection-pillars-2026) covers those in detail. **See what a deepfake document looks like under forensic analysis** — Upload a real document - or a fake one - and get a plain-English verdict in about a minute. $5 in free credits, no contract. (https://tampercheck.ai) ## FAQ ### Is a deepfake document the same as a tampered document? No. A tampered document started out genuine and had specific details edited - a balance, a name, a date. A deepfake document never existed in any real form; it was generated entirely by AI, branding and all. Both can look convincing, but they leave different forensic traces, which is why TamperCheck checks for both rather than assuming one implies the other. ### Can a deepfake document really fool a human reviewer? Yes, routinely. High-quality AI-generated documents are designed to satisfy a visual inspection - correct fonts, correct branding, plausible numbers. Studies on AI-generated fraud submissions put manual-review pass rates above 80%. The signals that give a deepfake away are statistical and structural, not visible to the eye. ### What industries are most exposed to deepfake documents? Anywhere a document is the basis for a decision: lending and underwriting (fake bank statements and payslips), KYC and onboarding (fake IDs), insurance (fake invoices and claims documentation), and education and hiring (fake transcripts and certificates). See [Synthetic Identity Fraud: Why the Document Layer Is Where You Stop It](https://tampercheck.ai/blog/synthetic-identity-fraud-document-verification) for how this plays out specifically in identity-based onboarding. ### How do I check if a specific document is a deepfake? Run it through TamperCheck via the [dashboard](https://tampercheck.ai/auth/login?returnTo=/dashboard) or [API](https://tampercheck.ai/docs). It runs 130+ forensic checks, including a dedicated AI-generation and deepfake detection layer, and returns a plain-English AUTHENTIC or TAMPERED verdict with the specific findings in about a minute. --- ## Where to go next - [What Is a Deepfake Document?](https://tampercheck.ai/problems/what-is-a-deepfake-document) - the full forensic breakdown on tampercheck.ai - [Detect Deepfake IDs](https://tampercheck.ai/problems/detect-deepfake-id) - the identity-document-specific version of this problem - [AI KYC](https://tampercheck.ai/ai-kyc) - the forensic layer that catches deepfakes inside an existing KYC stack - [Deepfake Document Fraud: How AI-Generated IDs Are Slipping Through KYC Checks](https://tampercheck.ai/blog/deepfake-document-fraud-kyc) - a deeper look at the KYC-specific attack patterns - [Run a free tamper check](https://tampercheck.ai/sample-report) - see a sample forensic verdict