AI-Generated Fake ID Documents Are Defeating Identity Checks

TL;DR: Identity verification vendors report steep growth in AI-generated passport and licence images and in manipulated verification video. Every figure counts the vendor's own traffic, and one vendor states that its own detection improvements inflate part of the increase. NIST's federal standard says live capture and liveness detection are "not sufficient to address all possible cases" of the attack behind it. The response is more identity collection, and the collectors keep getting breached.

A Different Problem From Synthetic Identities

This is not about fabricated people who build real credit over years. That pattern, a real Social Security number paired with an invented name and address, is covered in Synthetic Identity Fraud: AI Fake People, Real Loans.

The subject here is narrower and faster: a generated or edited image of a passport or driver licence, and a selfie video that has been swapped or replaced, used to pass one verification event. One account opened. One Know Your Customer check cleared. No fictional person required, because a real person's stolen documents will do.

What the Published Numbers Actually Measure

Three verification companies published data on this in 2025 and 2026. All three sell detection or verification services, all three counted their own customer traffic, and none measured an industry.

Resistant AI, a document-fraud detection vendor, reports "a 90x increase in Gen AI detections between 2025 and 2026" across more than 170 million documents processed for its own customers, and its own report attributes that number to both improved detection capability and real growth in the threat ^[1]. An unknown share of the 90x is therefore the vendor getting better at seeing fraud that was already there. A smaller figure sits beside it: roughly 5 percent of fraud declines in its adaptive decisioning sample are attributable to AI-generated documents, and most of the 2025 increase came after a November 2025 image-generation model release ^[1].

AU10TIX, an identity verification vendor, reports a 3.89 percent confirmed fraud rate for the first quarter of 2026, about 1 in 26 verification requests. That is confirmed fraud inside the payments, banking and trading traffic of AU10TIX customers, not an industry rate: payments 5.37 percent, banking 2.11 percent, trading 0.95 percent, with passports highest by document type at 7.89 percent ^[2]. Chief executive Yair Tal: "Financial services organizations are no longer defending primarily against altered physical documents." ^[2]

Sumsub, also a verification vendor, supplies the corrective. Its own platform's detected identity fraud rate fell from 2.6 percent in 2024 to 2.2 percent in 2025, still above the 2.0 percent of 2023 ^[4]. AI-assisted forgery rose from zero to 2 percent of the fake documents Sumsub detected in 2025, over what Sumsub notes is a six-month period only ^[5]. Its much-repeated 311 percent figure is real but tightly scoped: synthetic identity document fraud in North America, first quarter of 2025 against the same quarter of 2024, in a Sumsub-defined category on Sumsub's own platform ^[3].

Two more circulating figures need labels. Deloitte's 40 billion dollars in US gen-AI-enabled fraud losses by 2027 is a projection its Center for Financial Services modelled in 2024, by assigning risk scores and growth assumptions to fraud categories and anchoring to 12.3 billion dollars in 2023 ^[6]. Nobody has measured it. The FBI's Internet Crime Complaint Center logged 22,364 AI-related complaints in 2025 with 893,346,472 dollars in adjusted losses, but the AI section of that report names business email compromise, romance and distress scams, employment scams involving voice spoofing, and investment scams ^[7]. Document forgery is not among them.

Injection Attacks Are Not Presentation Attacks

The clearest description of the mechanism is in no vendor report. It is in NIST Special Publication 800-63A-4, the July 2025 federal standard for identity proofing that government and much of the private sector build remote verification against ^[8].

NIST separates two things marketing tends to blur. A presentation attack shows something false to the camera, and NIST requires presentation attack detection meeting an impostor attack presentation accept rate below 0.07 and conforming to ISO/IEC 30107-3:2023. An injection attack does not use the camera at all: "Injection attacks insert modified or forged media between the capture point (e.g., a device) and the element conducting the comparison or other operation (e.g., a server running the algorithms, a workstation used by a proofing agent)." ^[8]

NIST is direct about the cost to the defender. "A biometric comparison performed with a captured sample does not prevent these attacks." Live document capture and presentation attack detection "do provide some protection from injection and forged media attacks by making the injection of viable forged media more challenging," but "even these mechanisms are not sufficient to address all possible cases of these kinds of attacks." ^[8]

That is the sentence to hold onto when a platform explains that its new selfie check has closed the hole.

What the US Regulator Has Said

FinCEN, the Treasury bureau that collects suspicious activity reports from US banks, issued Alert FIN-2024-Alert004 on 13 November 2024 ^[9]. Its text: "Beginning in 2023 and continuing in 2024, FinCEN has observed an increase in suspicious activity reporting by financial institutions describing the suspected use of deepfake media in fraud schemes targeting their institutions and customers. These schemes often involve criminals altering or creating fraudulent identity documents to circumvent identity verification and authentication methods." ^[10]

That is a count of filings rising, not a fraud rate and not a loss total, and it moves with bank awareness as much as with crime. The alert's red flags are written for compliance teams: a customer photo that is internally inconsistent or inconsistent with other identifying information, multiple identity documents that do not match each other, geographic or device data conflicting with the documents submitted, and new accounts showing rapid transactions or high volumes of chargebacks and rejected payments ^[10].

Two Cases on the Record

BleepingComputer reported on 27 February 2026 that a 27-year-old Ukrainian national pleaded guilty in a Southern District of New York prosecution over a subscription platform that generated and sold more than 10,000 photos of fake identification documents, covering US driver licences, passports, Social Security cards and documents for roughly 56 other countries. Per that reporting of the indictment, "New York federal prosecutors said that the primary use of these fake digital documents was to circumvent Know Your Customer (KYC) verification requirements at banks and cryptocurrency exchanges." ^[11]

DutchNews.nl reported on 18 March 2026 that a man went on trial before the Amsterdam district court accused of opening 46 bank accounts in other people's names at ABN AMRO using doctored images of his own face against the bank's identity check, which asks applicants to submit photo ID and then verify it with a selfie. The pattern surfaced when one application paired a woman's photo ID with a selfie showing a man's face ^[12]. No verdict is reflected in the reporting cited here.

No fabricated person existed anywhere in the Amsterdam case. Real documents, a faked verification video, and one check standing in the way.

The Loop: Verification Creates the Material

Every response points the same direction, toward more document collection, more biometric capture and more centralised verification behind consumer apps. That infrastructure is itself a target, and it has already failed.

404 Media reported in June 2024 that AU10TIX, the vendor whose fraud data appears above, left administrative credentials exposed online for more than a year on a service that verifies identities for platforms including TikTok, Uber and X by processing photographs of faces and driver licences ^[13]. EFF, writing the same day, described the exposed logging platform as holding names, dates of birth, nationality, identification numbers and images of the documents themselves: "Hackers will just have to be lucky once." ^[14] The vendor background sits in Au10tix: Israeli Intelligence Ties, Data Breaches, and Your ID ^[20].

It kept happening. NPR reported in August 2025 that a women's safety app was breached, exposing driver licences, direct messages and selfies, government IDs that had been used as a verification tool, leaked onto a public message board ^[15]. In November 2025, researchers found an unprotected database traced to an identity verification firm holding more than 3 billion records, roughly 1 billion of them personal information across at least 26 countries, secured the day after notification with no malicious access confirmed ^[16]. Discord's mandatory age verification arrived four months after hackers took at least 70,000 government ID images from one of its verification vendors ^[21]. Twenty-five US states now require a licence upload or a face scan to reach certain sites ^[22].

The circuit closes there. Document forgery is why verification gets stricter. Stricter verification means more stored passports, licences and selfies. Those stores get breached, and breached document images are raw material for the next round. EFF's shortest version of the trade: "age verification systems are surveillance systems." ^[14]

If Your Identity Documents Were Exposed

Official guidance is jurisdiction-specific, and none of it fully fits a scanned copy leaking from a company rather than a wallet being stolen. A leaked copy cannot be reported lost, and replacing the card does not delete it. Reissuance, so the old number no longer validates, is the closest available remedy.

United States. USAGov points to the issuing authority for each document: the state motor vehicle agency for a driver licence or state ID card, the Social Security Administration website for a Social Security card, and the State Department for a lost or stolen passport, replaced in person. Most issuers require an official document such as a birth certificate to prove identity or citizenship ^[17]. Social Security numbers are not routinely reissued, and that page does not address the gap. On the financial side, the FTC says to call the companies where fraud occurred and ask them to close or freeze the accounts, then place a free one-year fraud alert with any one of the three credit bureaus, which must tell the other two, and pull credit reports even with a freeze in place. Reports go to IdentityTheft.gov, or RobodeIdentidad.gov in Spanish, which returns a free personal recovery plan ^[18].

United Kingdom. The ICO lists lost or stolen passports and driving licences as a warning sign of identity theft and gives a three-step chain: report the documents to the organisation that issued them, report the theft and any suspicious credit applications to the police and ask for a crime reference number, then contact CIFAS, the UK's fraud prevention service, to apply for protective registration ^[19].

What to Watch

Whether vendors separate injection figures from presentation figures. NIST draws the line, and the two attack types have different countermeasures ^[8]. A growth percentage that collapses both cannot be checked against the standard, and the collapsed number is always the larger one.

Whether detection improvement gets separated from threat growth. Resistant AI stated the confound in its own report ^[1], and Sumsub's detected fraud rate fell over the same period in which its AI-forgery share rose from zero to 2 percent ^[4]^[5]. A multiple published without that distinction is a marketing figure.

The two prosecutions. Judgment in the Amsterdam case was expected on 31 March 2026, and sentencing in the US case was scheduled for June 2026 ^[11]^[12]. Neither outcome appears in the reporting cited here.

The next verification-vendor breach. Four are already on the record above. The one that matters will be stored document images taken from a mandatory age or identity verification scheme, the point at which the fraud problem and the surveillance problem stop being separable ^[22].