Bill to ban creation of artificial superintelligence tabled at Westminster

Tabling is the cheap part of legislating. The record does not say what Alex Sobel's bill would prohibit, how it would define artificial superintelligence, or whether ministers will give it time ^[1]. Control AI, which is backing it, convened Monday's Westminster session, and the group is funded by Jaan Tallinn, an early investor in both Anthropic and Google DeepMind ^[2][15]. A backbench bill fronted by a lobby group is not a compliance horizon on this week's timeline. On the decade's, the definitions drafted now are the ones later drafters inherit.

The part of the week that already touches operations is the pre-release channel. According to the Financial Times, UK government officials were concerned that Anthropic had declined to submit its latest model, Mythos 5.1, to the AI Security Institute for pre-release testing, and the Cabinet Office confirmed that only a few US organisations have had access to it ^[10][11]. The tradeoff is not obscure: sending a frontier model to a national institute before launch costs schedule time and hands a government detailed capability information early, while declining protects both and spends the argument that voluntary arrangements are sufficient. The Cabinet Office still describes the institute as collaborating closely with industry partners including Anthropic ^[11], and Anthropic has been approached for comment ^[9], so what sits on the record is the decline rather than the reasoning behind it.

The figures being quoted at legislators are stated beliefs, not measurements. Anthropic's alignment science lead, Evan Hubinger, put the chance the technology could kill all humans in the next decade above 10% and said the company did not have a plan to ensure an aligned superintelligence ^[5]. Tallinn's estimate is that 10-15% of AI employees regard the technology as a worthy successor to humanity ^[16]. Bernie Sanders cites polling that 81% of Americans want their politicians to act ^[14]. A statute needs a threshold an auditor can test a model against, and none of those three quantities is that, which is the hardest part of drafting a capability rule.

Both labs' public positioning points at coordination rather than limits. OpenAI pointed to its chief scientist Jakub Pachocki's statement that international coordination on future AI development should become a top priority for governments ^[9], and Jacob Coxon, resigning from Anthropic with the claim that neither it nor OpenAI was acting responsibly, still said he was optimistic that US labs could coordinate on pacing ^[7][8]. Coordination leaves capability decisions inside the labs, while a threshold would move them out.

The same week produced three different instruments, and they are not interchangeable ^[17]. Sobel's is a prohibition ^[1]. Darren Jones wrote to the prime minister, Andy Burnham and the heads of the UN and the OECD asking for a multinational treaty and said explicitly that he wanted a safety-first approach rather than a ban on innovation or scientific endeavour ^[12]. Sanders wants Congress to legislate ^[14]. Jones's framing is the one that travels: whether the argument is about existential risk or dismissed as marketing hype, governments step in either way ^[13]. For a frontier lab this quarter, the live question is whether the voluntary testing channel holds, since there is no ban yet to comply with, and the best argument against a statutory threshold is a voluntary regime that visibly works, and the Mythos 5.1 record currently reads the other way.