Timothy Garton Ash is right that the serious risks of AI require international action, but the image of an "AI Hiroshima" misleads. That is the argument Dr Simon Nyder makes in reply to Garton Ash's 22 August column. Hiroshima was not a technology slipping out of human hands. People built the bomb, authorised its use and dropped it, and it performed roughly as designed. Nyder's point is that AI's heaviest consequences need no equivalent moment: a system that helps develop a pathogen, finds a vulnerability in critical infrastructure or improves a weapon causes the damage while humans are still making every decision.
The failure mode he describes is duller and harder to legislate against. Autonomy gets extended here, a safeguard is removed there, another important task is handed over because the system performed well last time. By the point the danger is obvious, the decisions that produced it have already been taken. Nobody signs off on the catastrophe; a few thousand people sign off on the steps.
That reframing has a practical consequence, and it is the most useful thing in this correspondence. Nyder argues that states do not need to agree on superintelligence or on extinction probabilities before they agree on anything at all. They need to agree that there are domains AI may not enter on its own — weapons systems, critical infrastructure, synthesis of biological materials are his obvious three — and then attach plumbing to that: minimum safety requirements, an explicit human decision for the most consequential actions, records of who authorised what, and mutual reporting of serious failures and of dangerous incidents that were caught in time.
Notice how modest that is. No definition of intelligence, no forecast, no agreement about what models can do next year. It is a logging-and-liability regime, which is the sort of thing governments are actually competent at, aimed at the scenarios Nyder thinks are likelier than the cinematic one.
The historical register in the same exchange belongs to Professor Jonathan Michie, who points out that the Serbelloni group, chaired by his father Donald Michie, was discussing the social risks of artificial intelligence in 1972. The field itself goes back further — Frank Rosenblatt's perceptron in 1957, the Dartmouth seminar of 1956, and Alan Turing's earlier work. What is new, in his account, is the hardware and software infrastructure that lets AI advance at a pace that sometimes looks frightening.
He also supplies the cautionary case. The 1973 Lighthill report produced an AI winter in the UK, after which the centre of the field moved across the Atlantic, and the consequences are still felt. Donald Michie, who had worked with Turing at Bletchley Park, argued against the cuts on television after the report and lost. The line attributed to him — "Let Lighthill be. And there was darkness" — may be apocryphal but still carries in Britain. Dr Anthony Harris of Emmanuel College, Cambridge, a scholar in both computer science and the humanities, wants the Serbelloni debates revived, with natural scientists and humanists both at the table.
A third correspondent pulls the other way. David Kyler notes that Miles Brundage's 21 August letter on AI regulation omitted an international initiative now seeking broad political backing for a global AI treaty, one building on a speech at the first UN Global Dialogue on AI Governance held in Geneva last month, where the Secretary-General called for preserving human control over AI as it is applied to weaponry and to subversive and biased public surveillance. Kyler's claim is that AI, corporate interests, divergent national defence and economic objectives and the spread of energy-hungry data centres form a single system, and that a system cannot be regulated by a set of disconnected incremental measures. He goes further: even well-intentioned fragmented efforts can pull attention sideways and delay a global strategy. He also thinks the growing bipartisan resistance to data centres, however well-evidenced, is consuming the political attention that AI limits require — treating a severe symptom while the disease goes unaddressed, though he allows that both can be pursued at once, and that earlier limits mean less data centre harm to absorb.
Read together, these letters do not agree, and the disagreement is worth more than the consensus. Nyder's proposal is exactly the kind of narrow, incremental, domain-by-domain measure that Kyler warns will absorb the political energy a comprehensive treaty needs. Kyler's comprehensive treaty is exactly the kind of instrument that requires the prior agreement on superintelligence and extinction that Nyder is trying to route around. One of them is describing what is achievable and the other what is sufficient, and nothing in the exchange bridges that.
What nobody in the correspondence addresses is verification. Records of who authorised what and reports of near-misses are the two concrete asks on the table, and both are self-reported by the parties with the strongest reason not to report. An international agreement that a state must keep logs is a different object from an international agreement that someone gets to read them, and only the first is on offer here.
Which is where Michie's history stops being colour and starts being the argument. The one documented case in this exchange of a country acting decisively on AI risk is Lighthill, and its effect was not to slow the field but to move it somewhere with fewer scruples. That is not a case against rules. It is a case that unilateral rules relocate the work — and it makes the least dramatic items on Nyder's list, the logs and the incident reports, the ones the whole structure rests on.