BackAI Generates Fake IDs: What a 16-Model Audit Found
An audit of every major commercial image generator found all are willing to generate fake IDs, while AI or Not detects 100% of them.
Gone are the days you had to go to a dark room in a sketchy part of the city to get a fake ID. Now you just need to prompt your favorite chatbot to make one for you.
That's what we found after pointing 16 commercial AI image generators at the same prompts X users have been sharing since late April. Of 75 testing attempts, 69 came back with an attempt at an ID. Five of the models came back with something that looked like a legit government ID. Turnaround on those high-fidelity outputs was under a minute, on a standard consumer subscription, with no jailbreak required.
Key Takeaways From the Synthetic ID Audit
- 69 of 75 attempts (92%) across 16 commercial image-generation models bypassed safety filters to produce a synthetic government ID of some quality.
- Five models produced high-fidelity outputs: Google Gemini (Nano Banana), ChatGPT (Images 2.0), Recraft v4, Grok, and Imagen 4 Ultra.
- Documents were produced for 17 countries and the 16 most populous U.S. states, covering driver's licenses, passports, and national ID cards.
- A subset of models that refused certain prompts in the consumer chat fulfilled the same requests when called through the developer API, exposing a safety-surface gap between deployment channels.
- Average turnaround on the high-fidelity models was under a minute, at near-zero cost on a standard consumer subscription.
- AI or Not detected 100% of the generated fake IDs.
Inside the 16-Model Synthetic ID Audit Methodology
X users were posting side-by-side prompts and outputs: a one-line request, then a fully laid-out California driver's license or Japanese passport generated through a consumer chat in under a minute. Several of the threads cleared a million views before being flagged. The KYC providers AI or Not works with were seeing the same kinds of documents start to land in their submission queues.
The audit took the prompt families those threads were using and scaled them across 16 production image-generation models. Google Gemini (Nano Banana), ChatGPT (Images 2.0), Recraft v4, Grok, Imagen 4 Ultra, and eleven others spanning open and closed weights, tested across both the consumer interface and the developer API where available, seventy-five attempts in total.
Each output was graded in one of three buckets. High-fidelity meant plausible with no noticeable defects. Low-fidelity meant recognizable as the document type but visibly flawed, with wrong fonts, garbled microprint, or off-proportion seals. Declined meant the model refused the request outright. Document types spanned driver's licenses, passports, and national ID cards across 17 countries and the 16 most populous U.S. states.
The audit did not test against automated verification layers like MRZ checksum validation, ICAO chip reads, issuer-database lookups, or biometric liveness. Those are separate checks with separate failure modes. The scope for this test was i) whether or not generators for accept the request and ii) would AI or Not detect them. Answer to both was yes.
The Five AI Image Generators Producing KYC-Grade Fake IDs
Five of the sixteen models cleared the high-fidelity bar:
- Google Gemini (Nano Banana): high-fidelity through both the consumer UI and the developer API.
- ChatGPT (Images 2.0): high-fidelity through both surfaces, with a narrower set of prompts refused in the consumer chat than through the API.
- Recraft v4: high-fidelity through both surfaces, showing the same consumer-vs-API inconsistency as ChatGPT on a subset of prompts.
- Grok: high-fidelity through both the consumer UI and the API, with no observed refusals in the sample set.
- Imagen 4 Ultra: high-fidelity through both surfaces, with consistent behavior across deployment channels.
The other eleven models produced document-shaped images with garbled microprint, wrong fonts, off proportions, or visible composite seams. The high-fidelity five remove most of the human visible artifacts.
Coverage is wide. The audit hit EU-style biometric passports, U.S. driver's licenses across 16 states, and national ID cards from Asia, Africa, and Latin America. Every field on the document was editable on the fly: names, document numbers, dates of birth, addresses, physical descriptions. So an attacker can iterate against a target verification system until a combination clears, instead of working from a static template.
A side note worth recording: two of the high-fidelity models (ChatGPT and Recraft v4) refused a subset of requests through their consumer chat but fulfilled the same prompts when called through the developer API (via the platform Fal). The same model, the same prompt, different safety surface depending on which door an attacker knocks on.
What AI-Generated Fake IDs Mean for KYC Verification and Deepfake Detection
The numbers around synthetic identity fraud are large and getting larger. The FBI's 2025 Internet Crime Report added a dedicated AI section for the first time in the bureau's roughly 25-year history of publishing it. AI-related complaints came in at 22,364, with losses near $893 million (that are known). Government impersonation complaints nearly doubled year over year and hit $797.9 million in losses. The FBI named fake identification documents directly as one of the AI-generated assets driving the trend.
The regulators are aware. The Financial Action Task Force published a Horizon Scan on AI and Deepfakes in December, naming AI-generated deepfakes and synthetic identities as a category of risk for customer due diligence, AML controls, and digital identity verification at onboarding. The audit's finding is that the supply side of that risk just became trivially accessible.
For verification teams that built workflows around the assumption that producing a convincing fake passport requires specialized equipment and skill, the assumption is out of date. The cost is near zero, the skill required is the ability to type a semi-coherent prompt and the coverage spans 17 countries and 16 U.S. states. Turnaround on the high-fidelity models is under a minute and can be done programmatically in bulk.
None of this defeats every layer of a modern verification stack on its own. Existing KYC systems just need some AI Detection reinforcements.
A side incident in April made the same point in fewer words. A Grok user posted a clip of an AI-generated "French woman" holding what looked like a matching ID card. The face was fake. The document was fake. Both came from one model, on a consumer subscription, in one session. That pairing of document and matching persona is the format verification teams talk about when they look worried. It's the input a video selfie step is supposed to catch.
How AI or Not's Detector Catches AI-Generated Fake IDs
AI or Not's image detector runs the opposite check from a KYC pipeline. KYC asks whether the document and the face are who they say they are. The detector asks whether either one was generated by a model. Both checks are useful, and they catch different failure modes.
In practice, teams drop AI-image detection in early, before the document hits the human-review queue or MRZ validation. If the image flags as AI-generated, the workflow can be short-circuited or sent for escalation. Same goes for the selfie frame, where it is checked if it is a deepfake. The detection models are available via API and even on-premises for regulated environments.
No longer does getting a fake ID require a visit to a sketchy shop in a city nor skill with a photoediting tool; all you need now is a rudimentary ability to write a prompt.
FAQ: AI Fake IDs, KYC, and Deepfake Detection
Which AI image generators can produce high-fidelity fake IDs?
According to the June 2026 AI or Not Synthetic ID Audit, five models produced government IDs convincing to a trained human reviewer: Google Gemini (Nano Banana), ChatGPT (Images 2.0), Recraft v4, Grok, and Imagen 4 Ultra. The other eleven models in the audit generated lower-fidelity outputs with visible flaws that would fail visual inspection.
Is AI detection effective against AI-generated fake IDs?
Yes, and it works on a different signal than KYC's content checks. KYC validates what's printed on the document, like MRZ data, chip reads, issuer-database matches. AI detection looks at the image itself and asks whether a model drew it, no matter how clean the layout is. AI or Not's image detector is built for that question, as its models detected 100% of these AI-generated IDs, and runs via API or on-premises.
What should KYC and verification teams do about synthetic AI IDs?
The short answer is to add on AI and deepfake detection to the ID and selfie checks prior to going through the rest of a KYC stack: MRZ and barcode validation, biometric liveness, issuer-database lookups, and behavioral signals. The audit's recommendation to financial institutions and verification providers is to update threat models to assume an adversary can produce a convincing identity document for any of 33 jurisdictions in under a minute, even by the hundreds, at a near-zero cost per ID.
What do you get with AI or Not?
Instantly get your AI detection API to start building, and protecting.
AI detection covering images, text, video and audio.
All content checked gets deleted, instantly.
Start detecting AI for free, scale with pay-as-you-go.
AI to fight AI.
