Human rights committee asks for a UK AI bill that grades systems by risk level

The sentence that reaches a product team sits in the middle of the summary. Obligations should be placed on "all stages of the AI lifecycle, ensuring those responsible in the supply chain for designing, amending and utilising AI systems are aware of their obligations and they are effectively addressed," the committee says ^[6]. Read plainly, that covers the company that licenses somebody else's model, edits the system prompt and puts the result in front of UK customers.

The committee's diagnosis of the current rules is that they are "fragmented and difficult to navigate", "leaving gaps in protection" ^[7]. Its examples of harm already done are specific: AI used to create sexualised images of women and girls, and scanning people's faces without their consent ^[8]. Bias appears as a training-data problem, with models trained on large datasets scraped from the internet that can include racist, sexist and other undesirable material ^[9].

The summary leaves the audits, dataset records and model documentation a company would have to produce to the bill. BBC News said it was seeking comment from the government on the call for one ^[10].

"It is moving with such speed and complexity that its impact is hard to accurately predict. What is clear is that at present we are unprepared to deal with its consequences however potentially dire they may be," Sobel said ^[11].

The report followed by two days a proposal from Dario Amodei, head of Anthropic, covering global regulation, industry-wide regulation and independent monitoring of AI models as they are developed ^[12]. Amodei urged the pace of development to be slowed, and Sam Altman of OpenAI and Elon Musk said they agreed with him ^[13].

Two axes sort the AI features a company ships into the UK. The first is whether the feature touches biometric data, profiling or subliminal technique, which is where the committee's candidate prohibition list points ^[5]. The second is whether the team could produce this quarter a written account of what data trained the model, who approved the deployment, and what was tested before release. Features that sit on the wrong side of both axes are the ones a risk-tiered regime would reach first ^[4], and they are also the ones an internal incident review would struggle with tomorrow. Writing that account now costs engineering time spent ahead of any published bill. Set against that cost, the report was agreed by a committee of twelve members from both Houses, drawn from the Labour, Conservative and Liberal Democrat parties ^[14].