AINEWS
The watchdog's ask: binding 'safe by design' rules requiring AI developers to ship models structurally incapable of producing this imagery.
The Internet Watch Foundation assessed 6,310 AI-generated images in the first half of 2026 that met the UK legal definition of child sexual abuse material. That is 40% more than the 4,500-plus it logged for the whole of 2025. The organisation says the rise is structural, not cyclical, and is now calling for binding legislation on the AI model developers whose products generate the imagery in the first place.
The IWF is a UK-based non-governmental body that monitors child sexual abuse material online, with analysts who assess images against the legal definition in Britain and report to international hotlines and platforms. Its new first-half 2026 figures cover still images only; the group has previously flagged a faster rise in AI-generated video, which it tracks separately. The majority of the images it assessed in 2025 and 2026 are of girls aged 7 to 13.
The proposal the IWF is putting to the UK government is a "safe by design" duty on generative-AI developers: a requirement that models be built so they cannot produce this category of imagery at all. IWF head of policy Hannah Swirsky framed it in plain terms: "It is incumbent on tech companies to build tools which cannot be abused this way. This problem has been created by technology, and if companies won't build tools which are safe by design, then they need to be made to do so by binding legislation."
A safe-by-design duty is structurally different from the after-the-fact content moderation that platforms run today. Moderation removes what the model has already produced. A safe-by-design duty targets the model's architecture: training-data filtering, refusal tuning that holds against jailbreak attempts, and evaluations that demonstrate a model cannot generate CSAM under adversarial prompting. It is a duty on the developer's product, applied before deployment, not a duty on a host platform's review queue.
The UK already has a related offence. It is illegal to adapt an AI model to create child sexual abuse material, or to distribute such a model to others. Developing a model designed to produce hyper-realistic CSAM carries up to five years' imprisonment. A UK government spokesperson confirmed that existing law applies regardless of whether the material is AI-generated. The IWF's argument is that the act-based offence punishes the person who adapts or distributes; it does not require the upstream model developer to ship a model that is structurally incapable of producing the imagery in the first place.
The political window in London is open. AI minister Kanishka Narayan told the Guardian that "nothing is off the table" on further regulation and named the test: "Will it make the British people safer?" Andy Burnham's government has not signalled that a new AI bill is imminent, and Narayan's "nothing is off the table" line is the closest the government has come to confirming that one is being scoped.
A secondary indicator is the IWF's Report Remove service, which took 420 reports from children in the same period about images of themselves they believed had been faked or manipulated. That figure is reported separately from the 6,310, but it points in the same direction: the victims are finding the material themselves, often after it has already circulated.
The choice for the UK is whether to be the first jurisdiction to impose a duty on the model rather than the user. The data is on the table. The minister's test is on the record. The next step is whether a bill gets drafted, or whether the existing offence is held to be enough.