Best AI-Generated Document Detection Tools
Eight AI-generated document detection tools compared by fit, what each reads, and its limits, plus why no detector catches every forged file.
TL;DR: Fake documents cluster around four frauds: inflated income, invented funds or payments, companies that are not what they claim, and tax records the IRS never received. This list covers 20 file types, each with its strongest tell and a full guide. GAO found alleged tax document falsification in 58 percent of 330 pandemic loan fraud cases, yet Sphinx found no regulator dataset that ranks forgeries by document type, so the numbers below are for navigation, not rank.
What "Most-Forged" Means Here
"Most-forged" on this page is an editorial call, not a measurement. Sphinx looked for a regulator or government dataset that counts forgeries by document type and did not find one. Detection vendors publish network data by document type, but each set reflects that vendor's own customers rather than the market.
The closest public evidence comes from two studies. In GAO's review of 330 PPP and COVID-19 EIDL fraud cases brought by the Department of Justice, 227 cases (69 percent) involved falsification or alleged falsification of tax or other documents such as payroll records and bank statements, and 190 (58 percent) involved tax documents specifically. Those cases come from one federal program, announced by December 2021. The ACFE's Occupational Fraud 2026 report, drawn from 2,402 cases, found that 39 percent of fraudsters concealed their schemes by creating fraudulent physical documents, 33 percent by altering physical documents, and 28 percent each by creating or altering electronic files. That study covers fraud inside organizations and counts methods, not file types.
So each entry had to pass two tests: a financial institution relies on it to release money, open an account, or clear a business, and the applicant, not the issuer, controls the copy the reviewer receives. Public evidence is strongest for income, banking and tax files and thinner for corporate records, which earned a place because KYB controls depend on them. Numbers run in group order, and group order is not a ranking either. Identity documents are a separate problem, covered in identity document verification for customer onboarding.
Two Ways a File Goes Wrong
A fake document is either edited after creation or generated from scratch. An edited file began as something real and was changed later, so it tends to leave a mismatch around the changed field and a file history that continues after the issuer's system finished. A generated file never passed through the issuer, and template-made files have no issuance history to find. It can be internally perfect and still fail against any record outside the page.
FinCEN's November 2024 deepfake alert warned that GenAI can produce content that is difficult to distinguish from unmodified outputs. The FBI's April 2026 release on its Internet Crime Report recorded 22,364 AI-related complaints costing nearly $893 million, naming fake identification documents among the tactics. That is a broad fraud figure, not a rate for these files.
Fake Income Documents: Earning More on Paper
Income documents are fake financial documents built to change one number: what a person can afford. They often arrive as a set, so test them against something the applicant does not control.
1. Pay stub
A payroll PDF altered after issue or generated outright to inflate income for a loan or onboarding.
Strongest tell: net pay that never appears as a matching bank deposit. Generated stubs increasingly get the arithmetic right, so math alone is only a screen. How to spot a fake pay stub.
2. Employment verification letter
A letter claiming an employer confirmed a job, title, dates, or pay. Two applicant documents agreeing with each other proves nothing.
Strongest tell: the employer's own channel does not confirm it. A phone number printed on the letter does not count, because the applicant chose it. How to spot a fake employment verification letter.
3. Lease agreement
A contract edited after signing or assembled for a tenancy that never existed, offered as rental income in underwriting or as proof of address or premises. Leases have no standard format, so layout proves nothing.
Strongest tell: a landlord who exists nowhere but on the lease, with no registry entry and no link to the property. How to spot a fake lease agreement.
Fake Banking Documents: Money and Payments That Never Existed
Banking documents borrow an institution's credibility. Each can be checked against that institution or the money itself. For how banks layer detection, see document fraud in banking.
4. Bank statement
A PDF or image misrepresenting balances, income, or activity. Looking like a bank PDF is no longer a finding.
Strongest tell: a running balance that does not follow from the line above, or a file history that continues after the bank's system produced it. How to spot a fake bank statement.
5. Proof of funds letter
A bank letter edited to change a balance, date, or name, or built on lifted letterhead for an account that may not exist.
Strongest tell: a mismatch with how that bank actually issues such confirmations, such as a branch manager signing from a personal email. How to spot a fake proof of funds letter.
6. Bank reference letter
A page claiming a bank attested to a customer relationship it never attested to, or attested to differently.
Strongest tell: language a real reference does not use. A genuine status reply hedges, while a page that warrants an unencumbered balance has left the document type. How to spot a fake bank reference letter.
7. Wire confirmation
A payment advice or portal receipt edited or generated to describe a transfer that never settled.
Strongest tell: amount, beneficiary, and reference that describe a different payment from the contract or invoice behind it. How to spot a fake wire confirmation.
8. Voided check
A routing and account number presented as a bank document, which then becomes a payout instruction for payroll, vendors, or marketplace sellers. A genuine check can still belong to someone else.
Strongest tell: no evidence the applicant controls the account. A VOID mark and a logo say nothing about ownership. How to spot a fake voided check.
9. Brokerage statement
A statement offered as proof of assets that misstates positions, cash, or market value.
Strongest tell: a stated total that cannot be derived from the positions and cash on the same page. How to spot a fake brokerage statement.
Fake Corporate Documents: A Company That Only Exists on Paper
Fake business documents do their damage at KYB and in business credit. Formation and status records have a public issuer to check against. Internal records usually do not, so they get tested against the rest of the file.
10. Certificate of incorporation
A formation filing generated from scratch, edited after a real export, or reused across unrelated applicants.
Strongest tell: registrar mismatch. The number on the page must resolve at the named registrar to the same name and date. How to spot a fake certificate of incorporation.
11. Certificate of good standing
A status PDF generated, edited after a real download, or swapped for a portal screenshot, so an entity looks authorized on a date the register does not support.
Strongest tell: the register disagrees, on the entity or the date. The office named on the page is the office that can confirm the status. How to spot a fake certificate of good standing.
12. Operating agreement
An LLC's internal contract on members, ownership, and signers. Real ones are often templates, so a form-like layout means little.
Strongest tell: internal disagreement, such as percentages that do not sum or an effective date before the LLC was formed. How to spot a fake operating agreement.
13. Board resolution
A certified copy of bank authority the board supposedly adopted, naming who may open accounts and sign. An unfiled resolution is normal, so a missing state seal is not a finding.
Strongest tell: a certifier who is the only signer the authority names, or a meeting dated after the account was already opened. How to spot a fake board resolution.
14. Financial statements
A P&L, balance sheet, or audited report edited after export or generated to describe performance that never happened.
Strongest tell: figures that do not tie to records the borrower does not control, such as bank deposits. How to spot fake financial statements.
15. Invoice
A real invoice edited after issue, or a generated one describing a sale that never happened. It works as factoring collateral and as a payment instruction in accounts payable.
Strongest tell: a remittance account that differs from every prior invoice from that vendor. How to spot a fake invoice.
The Rest of the Series
Eleven more guides sit outside the numbered list, because they serve a narrower workflow or fall outside the four groups: utility bills, for instance, are proof of address.
The Page Is a Claim. The Record Is the Answer.
Almost every tell above connects the document to something the applicant does not control: a deposit, a register, a transcript, an issuer's own phone line. Logos, seals, fonts, and letterhead appear on the list only as things that fail.
File-level evidence is the second line. An edit often leaves a production and timestamp trail the page does not show, which PDF metadata analysis covers in depth. Neither line is complete alone: a clean file history does not make a genuine check the applicant's own. Document verification and document fraud detection answer different questions, and a serious program asks both. For rates rather than tells, read the document fraud statistics and check what each figure measures.
What to Do With a Suspect File
Treat a tell as a reason to verify, not a verdict. Genuine files have odd fonts and rounded numbers, so one anomaly opens a review rather than closing the case.
- Keep the original file as received. A screenshot or re-saved copy loses its history, so ask for the download the issuer produced.
- Run the entry's external check using a contact or registry you sourced yourself, never one printed on the document.
- Read the packet together. Files that agree with each other prove little if they all came from the applicant.
- Record what you checked, what it showed, and when.
- Escalate under your policy. Whether a SAR is warranted belongs to your BSA team, and the applicant should not hear which check failed.
For a second read on a single file, check a document in the free playground.
Where Sphinx Fits
Sphinx Doc Fraud examines the file itself through six checks: production method, timestamp trail, issuer matching, consistency, model artifacts, and recycled patterns. Sphinx reports a 94.3 percent correct verdict and under 28 seconds to clear clean files. Pricing is $0.45 per document, with no seats and no platform fee, and the Doc Fraud playground is free to try.
The limit is worth stating. Doc Fraud reads the file. It does not verify a person's face or connect to a bank account, so the deposit, registry, and transcript checks above stay in the process. Inside a KYB or KYC workflow it runs as one agent among several, where Sphinx prepares the evidence, applies your SOPs, and produces a traceable recommendation. Your team retains final approval.
Frequently Asked Questions
What is the most commonly forged financial document?
No regulator dataset Sphinx found answers this by document type. GAO's review of 330 PPP and COVID-19 EIDL fraud cases points to tax documents, alleged in 58 percent of them, but that covers one federal program. Vendor figures reflect each vendor's own customers.
How can you tell if a document is fake?
Test the document against a record the applicant does not control: a deposit, a registry entry, an IRS transcript, or the issuer reached through a contact you sourced yourself. Then look at file history for signs of editing after creation. Fonts and logos only catch crude work.
What is the difference between fake financial documents and fake business documents?
Fake financial documents misstate money, such as income, balances, or payments, and usually support a loan or a payout. Fake business documents misstate a company, such as its existence, status, owners, or signers, and usually appear at KYB. Financial statements and invoices sit in both camps.
Can AI-generated fake documents be detected?
Yes, though rarely by eye. FinCEN warned in November 2024 that GenAI output can be difficult to distinguish from unmodified content. Detection relies on production method, model artifacts, and recycled patterns in the file, plus checks against outside records that a generated page cannot satisfy.
