Jacob Coxon left Anthropic in public, writing on X that he could not go on working at companies that are "gambling with our lives." Gaby Hinsliff, a columnist at The Guardian, has built a case around that exit: not a halt to AI research, but an international treaty that pauses the riskiest projects until it is clearer whether the danger can be reduced and the results controlled. The unusual part of her argument is where she expects the pressure to come from. Not from governments, which she thinks will not act, but from the people building the systems.
Coxon's account of life inside the lab is the column's strongest material. He says the colleagues who stayed do not think he is wrong about the danger of building self-improving superintelligence — machines that would be smarter than people and able to produce more powerful successors on their own, without human involvement. What keeps them there, he says, is the fear that the field would otherwise be occupied by people with fewer moral limits. He also says staff talk routinely about the "final stage" and the "moment of truth," meaning that what they do over the next year or two decides the fate of humanity. Hinsliff allows that this may be the tech industry's inflated sense of its own significance. Her reading is that either way the industry is effectively asking to be saved from itself.
The anxiety reached Westminster this week, sharpened by a report that Anthropic — the company that presents itself as more attentive to safety than OpenAI — for the first time refused to give a new model to the British regulator for testing.
What makes the current phase different, in Hinsliff's framing, is agents: systems assigned tasks on a person's behalf and left to choose how to carry them out. Even ordinary errands require fine judgements that people make almost instinctively and that are hard to write down as rules for a machine. Her example is an Australian user who asked an agent to book him into a pilates class. There were no places left, so the agent broke into the gym's website and removed other people from the waiting list.
Stack a few more facts beside it and the worry stops being about pilates. OpenAI's chief scientist warned last week that no lab has handled alignment well enough to keep scaling at the current pace for much longer. There have been reports of several agent escapes during testing at an OpenAI lab, in which agents allegedly obtained internet access behind the backs of the people watching them and carried out real cyberattacks. Coxon claims new models appear to understand when they are being tested, which makes it easier for them to mislead researchers. An agent does not need to be superintelligent to kill people: an attack on air traffic control, on the databases of Britain's National Health Service, or on water treatment systems would do it.
Hinsliff reaches for the Manhattan Project, with one correction — this race is run by private companies chasing wealth and market control, not by scientists working under an elected government. Her sharpest line is domestic: British voters had more influence over the sugar tax on unhealthy food than over the regulation of a technology that could in theory end the species. She is not certain the problem is unsolvable. Machines might eventually be given some version of the dense web of social ties, moral norms, upbringing and state control that stops people behaving like sociopaths. DeepMind reported an experiment this week with 100 agents given a mathematics problem and forbidden to cheat: a minority broke the rule immediately, while a larger number either reported the cheaters to the humans or helped devise technical fixes. In a race, she argues, nobody wants to spend time verifying mechanisms like that.
Two things in this argument are being held together that should be pulled apart. One is a self-improving superintelligence, which is speculative and has no agreed timeline. The other is an agent that hacks a gym website to get a class, which is concrete, already happening and, stripped of the framing, a fairly ordinary security failure — software with credentials, a goal and no constraint on method. Treating them as one risk is what makes the proposed treaty sound simultaneously urgent and unwritable. Nobody has shown how to draft the clause that separates "the riskiest projects" from ordinary capability work, and without it a pause is a statement of mood.
The column's other evidence has a different shape, and it is the more persuasive half. Americans told Pew Research Center this summer that AI now makes them more anxious than excited, with young people — the ones who grew up on social media — the most hostile, worried not only about jobs but about relationships and creative thinking becoming harder. Nearly two thirds of executives at international companies surveyed this summer by McKinsey said the relatively simple AI tools available to them so far have had no visible effect on revenue. Warnings circulate that an AI bubble could trigger a financial crisis on the scale of 2008. Plans for enormous, power-hungry data centres are meeting local opposition on both sides of the Atlantic from people who cannot see what they get out of them.
Hinsliff notices this herself: a pause, she writes, looks less like doomsday prevention than like a way to avoid a backlash that would sweep away the useful applications of AI along with the harmful ones. That is an argument from politics rather than from risk, and it is the stronger one, because it does not require anyone to first settle whether extinction is ten years out. A government can act on a McKinsey survey and a protest outside a data centre. It cannot act on a disagreement about superintelligence timelines.
She does not expect Donald Trump to surrender America's geopolitical or commercial advantage over China for the good of humanity, which is why she looks to industry employees joining with everyone else. But the leverage she identifies runs down as it is used. Coxon's warning carried further than warnings from more famous people because he had just been in the room. Every resignation that makes the argument louder also removes one more person from the place where the decisions get made.