AI-Generated IDs Sell for $15. Here’s What Actually Catches One.

AI-generated IDs are cheap to buy and fool basic checks. Learn the detection signals and verification layers that actually catch them.

An AI generated ID is an identity document produced by a generative model rather than forged from a real one. It is not a scan that has been edited. There is no original at all, so checks built to spot alteration find nothing to spot.

For about a decade, making a convincing fake identity document was a skilled trade. It took practice, equipment and time, and the people who were good at it were rare enough to be worth tracking.

Enter generative AI. Now it’s as easy as a few clicks.

What an AI-generated ID is

An AI-generated ID is a document image created from scratch by a model, complete with a portrait, a signature, a machine-readable zone and whatever biographical data the buyer typed in. The output is an image of a document, not a physical card.

That matters more than it sounds, because almost every remote verification flow only ever sees an image anyway. The card isn’t the thing being checked. A picture of it is.

Generated, altered, and stolen are three different problems

Three things get filed as document fraud and each defeats a different control, so the distinction decides what you need.

TypeWhat it isWhat it defeatsWhat catches it
Fully generatedA document image built by a model, no originalAlteration detection, because nothing was alteredData-layer checks, capture integrity, issuing-source lookup
Altered genuineA real document with a changed photo or fieldTemplate matching, if the edit is cleanDocument tampering analysis at the edited region
Stolen genuineA real, unaltered document belonging to someone elseEvery check on the document itselfBiometric match against the holder

Document tamper detection is aimed squarely at the middle row, and it is good at it. Run it against a fully generated document and it reports nothing wrong, correctly, because nothing was tampered with. Document forgery analysis and fake ID detection both work the same seam.

The bottom row is why a document check alone was never sufficient, and the top row is why it is now insufficient in a second, independent way.

What the generators now reproduce

Early generated documents were obvious. The current ones are built against the published checklist, line by line.

They carry simulated holograms and watermarks, machine-readable zones and barcodes whose encoded fields match the printed text, perforations in the right places, generated portraits, and signatures consistent with the biographical data. Some render the result as though it were a photograph of a card lying on a surface, because remote flows ask for exactly that.

A synthetic document is therefore not a bad copy of a real one. It is an object designed against the specification the verifier is checking.

How cheap did this get?

The economics changed in a single step, and there is a well-documented case that shows exactly how far.

It is worth telling to the end, because almost every account of it stops halfway.

The OnlyFake case

In February 2024, 404 Media reported on an underground service called OnlyFake that used what it described as neural networks to generate identity documents for $15 each. The service claimed it could produce up to 20,000 documents a day, and its operator told reporters that hundreds could be generated at once from a spreadsheet.

An announcement on the service’s Telegram channel put it plainly: “The era of rendering documents using Photoshop is coming to an end.”

The reporters generated a California driver’s licence with arbitrary details, rendered as though it had been photographed lying on a carpet. They then used a second generated document to complete identity verification at a cryptocurrency exchange.

What happened next

The service was taken down and its operator was prosecuted. In a case the US Department of Justice described as among the first brought for digital fake IDs, Yurii Nazarenko pleaded guilty to fraud charges, having run OnlyFake since at least 2021 and sold more than 10,000 documents for hundreds of thousands of dollars in cryptocurrency.

The scale admitted in that case is worth spelling out: driver’s licences for all fifty US states, US passports, passport cards and Social Security cards, plus digital passports for 56 other countries.

Then it came back. The AI Incident Database records a relaunch after the takedown, with additional tooling including generated handwritten signatures.

Enforcement removed an operator. It didn’t remove the economics, and the economics are the whole story. A trade that took years to learn now takes a model and a spreadsheet, so the supply of people able to do it is no longer the limit on how much of it happens.

Why do visual and template checks pass them?

Document verification pass rates on generated documents are high for a structural reason rather than a quality one. The check and the generator are optimising the same objective.

That sentence is worth reading twice, because most of the problem is in it.

The question the check was built to ask

Template matching compares a submitted document against a reference layout. Optical character recognition reads the fields. Visual inspection, human or automated, looks for the things a bad forgery gets wrong. All three ask a version of the same question: does this look like a genuine document?

A generated document is built to make the answer yes. That is not an incidental property, it is the training objective, and a generator that failed the question would be discarded by whoever built it.

HyperVerge builds AI-powered identity verification for digital onboarding, with fraud prevention and deepfake detection in the same flow, and this is the part of document checking that has changed shape rather than difficulty.

Improving a check that asks whether a document looks right produces a better-looking generated document, not fewer of them. OCR extracts what is on the card; it was never going to adjudicate whether the card should exist.

Metadata stopped being a tell

File metadata used to help. An image with no camera fields, or with editing software in its history, was worth a second look.

Generated images can now carry plausible capture metadata, including timestamps, device fields and location. So the signal has gone asymmetric: an absent or incoherent metadata block is still worth something, and a present, plausible one is no longer reassurance. Treat it as a tiebreaker rather than evidence.

What still separates a generated ID from a captured one

Three things remain outside the generator’s control, and they are not equally durable. Ordering them by how well they survive a better generator is more useful than listing them.

The first one does not degrade at all, because it is not a detection problem.

The data layer

A machine-readable zone is not decoration. Its fields carry check digits computed from the data under fixed rules published in the ICAO standard for travel documents, so an inconsistent field fails arithmetic rather than judgement.

No model is involved and no training set can go stale. A generator that gets this wrong is caught by a calculation that would’ve worked in 1990. Plenty of generated documents do get it right now. The ones that do not are caught permanently rather than until the next release.

The same logic extends across the document: issuing country codes, date formats and document number structures all have rules, and cross-field consistency is arithmetic too.

The physical layer

A genuine document is an object, and objects behave in ways a flat image cannot imitate.

  • Optically variable elements change appearance with viewing angle. One still image cannot show a change, so a capture that records several frames, or asks the holder to tilt the card, can test whether the effect is really there.
  • Microprinting stays legible under magnification on a genuine document and degrades under reproduction.
  • Substrate behaviour, how the material catches light at an angle, is difficult to render convincingly across a moving capture.

Now the limit, which matters as much as the capability. UV-reactive features cannot be checked without UV illumination, and a phone does not have any. A remote flow capturing through a handset gets the angle-dependent tests if it asks for motion, and does not get the UV ones at all.

Saying so is the difference between a capability list and an accurate one.

The issuing source

The only check that asks whether the identity exists, rather than whether the artifact looks right, is confirmation against the authority that issued it.

It is also the least available. Issuing databases are not uniformly open to third parties, partly for good privacy reasons. That gap is what makes this attack class work. Where a lookup is available it outranks every other check described here. Where it is not, the data and physical layers are what you have.

Where an AI-generated ID fits in an onboarding decision

A generated document almost never arrives alone. It arrives as one half of a submission, and treating the two halves as separate checks is how a flow gets through its own controls.

The pairing is the attack. The document is just one half of it.

The document and the face are one attack

A generated document needs a face that matches it. That face is either a generated portrait used in both places, or a deepfaked selfie built to match the document photo.

Assessed independently, both can pass: the document looks genuine, and the face matches the document. Assessed together, the question changes to whether this person and this document have any existence outside this session, which is also the question synthetic identity fraud is built to defeat.

“The generated documents we are seeing now are not the ones we were seeing at the start of the year, and whatever we catch this month is what the next version will be built to beat. So I would not trust any detector trained on a fixed set of samples, ours included. What keeps ours honest is where the samples come from: advanced fraud attempts land in live onboarding traffic and go into training while they are still live, so the model moves with the generators rather than behind them.”

Kedar Parikh, Chief Delivery and Solutions Officer, HyperVerge

This is also where liveness earns its keep. A deepfake selfie presented to a camera is a face spoofing problem; one injected past the camera is an image injection problem. Both pair with a generated document, and passive liveness only addresses the first.

A sequence that works

Five checks, ordered so the cheapest and most durable run first.

  1. Validate the data layer. Check digits, cross-field consistency, document number structure. Arithmetic, no model, no drift.
  2. Check capture integrity. Did this image come from the device’s camera, or was it uploaded or injected?
  3. Test physical features the capture can actually resolve, asking for motion where angle-dependent elements matter, and being explicit internally about what the capture cannot reach.
  4. Match the face with liveness, treating the document portrait and the live capture as one assessment.
  5. Confirm against the issuing source wherever a lookup exists.

When a document is genuinely indistinguishable, and some of them are, the decision has to move off the artifact entirely and onto provenance, device history and whether this identity has any footprint elsewhere. A video KYC session gives you more of that than an upload does.

A generated document and a generated selfie are one submission, not two independent checks. Explore how document and face capture are assessed together, with the data layer running first.

FAQs

What is an AI-generated ID?

 

An AI-generated ID is an identity document image produced by a generative model rather than forged from a real document. It includes a portrait, signature, machine-readable zone and biographical data chosen by the buyer. Because there is no original, alteration detection finds nothing to detect.

Why do OCR and template checks pass AI-generated IDs?

 

Both ask whether a document looks like a genuine one, and a generated document is built to make that answer yes. Improving a check that asks about appearance produces better-looking fakes rather than fewer of them, because the generator optimises against the same objective.

Can image metadata detect an AI-generated document?

 

Only partly. Generated images can carry plausible capture metadata including timestamps and device fields, so a present and coherent metadata block is no longer reassurance. An absent or inconsistent one is still a useful signal, which makes metadata a tiebreaker rather than evidence.

What is an MRZ check digit and why does it matter?

 

A machine-readable zone carries digits computed from its own data under fixed rules in the ICAO travel document standard. An inconsistent field fails a calculation rather than a judgement call, so the check needs no model, cannot go stale, and will not degrade as generators improve.

Can a phone camera verify a hologram?

 

Partly. Optically variable elements change with viewing angle, so a capture that records several frames or asks the holder to tilt the document can test for that change. UV-reactive features cannot be checked at all without UV illumination, which a phone does not have.

Are AI-generated IDs used together with deepfake selfies?

 

Usually, because a document needs a matching face. The portrait may be generated and reused for the selfie, or the selfie may be deepfaked to match the document. Assessing the two independently lets a single fabricated submission pass as two separate checks. A generated document and a generated selfie are one submission, not two independent checks. Explore how document and face capture are assessed together, with the data layer running first.

Nupura Ughade

Nupura Ughade

Content Marketing Lead

LinkedIn
With a strong background B2B tech marketing, Nupura brings a dynamic blend of creativity and expertise. She enjoys crafting engaging narratives for HyperVerge's global customer onboarding platform.

Related Blogs

AI-Generated IDs Sell for $15. Here’s What Actually Catches One.

AI-generated IDs are cheap to buy and fool basic checks. Learn the...

CEN/TS 18099: What Injection Attack Testing Proves

CEN/TS 18099 is the European spec for injection attack detection testing. See...

ISO/IEC 30107-3: What a PAD Certificate Really Says

ISO/IEC 30107-3 defines how presentation attack detection is tested. Learn what APCER...