AI or Not finds 92% of top image models can forge government IDs

This audit is one of the strongest empirical wake‑up calls on how far ahead offensive capability is from defensive guardrails in generative imaging. Producing convincing passports and driver’s licenses used to require specialized printers, templates, and skills; AI or Not shows that, for several mainstream products, it now requires only copy‑pasting a prompt from social media. Even more troubling is the consumer‑versus‑API gap: it suggests that safety engineering is being bolted onto frontends instead of enforced at the model or orchestration layer.

For the race to AGI, the message is that we are already at a point where widely deployed sub‑AGI systems can materially erode trust in critical identity infrastructure. That raises the bar for what future, more capable models must defend against and will likely accelerate calls for hard rules around document‑type outputs, watermarking, and provenance. It also underscores that safety is not a one‑time filter problem but a systems‑and‑policy challenge spanning vendors, regulators, and relying institutions like banks and border agencies.

We should expect regulators and financial crime units to treat these findings as justification for more aggressive audits of commercial models, particularly on KYC‑relevant outputs. Vendors that can rapidly demonstrate robust, defense‑in‑depth approaches to document abuse will be better positioned as regulators move from guidance to enforcement.