An AI 'undressing' offender bailed opposite a playground: the tools moved faster than the paperwork

The facts here are thin, and we will not pad them. According to Sky News, a man convicted in connection with paedophilia who used artificial intelligence to digitally undress images of children was bailed to a flat that overlooks a playground. We have not seen the full case file, the court's reasoning or the housing authority's side, so we are commenting on what the report establishes and no more.

Even at that level, the story is a clean illustration of a mismatch we keep returning to. So-called "nudify" tools — image models fine-tuned or prompted to generate sexualised imagery from ordinary photographs — collapsed the cost of producing child sexual abuse material from something requiring access and effort to something requiring an app and a source picture. That capability arrived in roughly the span of a product cycle. The machinery that classifies an offender, scores their risk and decides where they may live moves on the timescale of statute, guidance documents and case law.

That is the thesis: the harm is now generated by a technology that is two or three years old, while the response runs on categories written for a pre-generative world. A risk framework built around physical contact and possession of files does not automatically know what to do with someone whose offending was a prompt. And a placement decision is only as good as the risk label feeding it — if the label is wrong or incomplete, no amount of downstream process catches a window facing a playground.

Our read: this is the transition cost, and it is being paid by the people least able to refuse it. We are long-term optimists about AI — we think it ends in cured diseases, longer healthy lives and abundance that frees people to work on what they love. We are not optimists about the next few years of institutional catch-up, and cases like this are why. The gap between what the tools can do and what our systems are configured to recognise is where the real damage lands, and it closes only if someone deliberately closes it.

The practical fix is unglamorous and does not require solving AI safety in the abstract: offence categories that explicitly cover synthetic abuse material, risk assessments that treat generative capability as a live factor rather than a technicality, and placement checks that actually cross-reference the nature of the offence against the location. None of that is a research problem. It is an administrative one, which is precisely why it gets left undone. Getting the boring layer right is how a society earns the good version of this technology instead of absorbing the worst of it first.