A Precedent for a Technology that could “Kill us All”

By William Leiss (15 September 2026)

A few weeks ago the AI researcher Jacob Coxon publicized his resignation letter in which he stated: “The people building AI earnestly believe that it could kill us all by the end of the decade.” He added: “This is not a marketing stunt.” Some of his erstwhile colleagues openly affirmed his view. Others outside the industry, such as the Director of the Harvard Center for Risk Analysis, remarked that it was not plausible. And yet the thought that AI could lead to human extinction has been repeated for years by none other that Geoffrey Hinton, the “godfather of AI” and winner of the 2024 Nobel Prize in Physics.

What should the rest of us think of all this? Would it help if we recalled an earlier episode in which a threat to the continuation of all human life was mooted?

During the Manhattan Project, a young physicist named Edward Teller, later known as the “father of the hydrogen bomb,” advanced a theory that the explosion of an atomic bomb could ignite a nuclear chain reaction in the atmosphere and also possibly in the oceans. Of course, this would mean the extinction of all we humans in addition to our animal cousins. When this supposition reached the ears of some higher officials outside Los Alamos, Robert Oppenheimer was told that the project might have to be put in abeyance. Hans Bethe, who was director of research reporting to Oppenheimer at the New Mexico facility, decided to commission a formal study of the idea. Teller and two colleagues were told to complete it, and fast.

Their short report, “Ignition of the Atmosphere with Nuclear Bombs,” was completed in early 1945. (It was publicly released in 1946 and can still be found online; I have a copy.) Most of it consists in equations, but the first page is in plain English. The opening sentences read: “It is shown that, whatever the temperature to which a section of the atmosphere may be heated, no self-propagating chain of nuclear reactions is likely to be started. The energy losses to radiation always overcompensate the gains due to the reactions. This is true even with rather extravagant assumptions concerning the reactivity of the nitrogen nuclei of the air.”

Note especially the conclusion that the feared event is not likely to occur. Perhaps a low risk, or even a very low risk? The authors do not say. The great Enrico Fermi, who was also at Los Alamos, and who had built at the University of Chicago (with Leo Szilard) the world’s first controlled nuclear reactor, was a notorious jokester. Just before Trinity, the first-ever detonation of a nuclear device, he was taking bets with his colleagues on whether the atmosphere would ignite. But on 16 July 1945 they pushed the button. And we are still here.

My suggested solution to our dilemma about AI is that we should do what Hans Bethe did some seventy years ago: Commission an authoritative study. And the USA has the perfect candidate for the job, namely, the National Academy of Sciences (NAS). The NAS and its peers elsewhere have a splendid track record in this regard, having published hundreds of careful, disinterested expert panel studies on often contentious topics. For greater effect, NAS could partner with the oldest of national academies, The Royal Society in the U.K., and also with its North American partner, The Royal Society of Canada. (I was once a member of a trilateral U.S.-Mexico-Canada panel study headed by the NAS, and it was a remarkable experience.)

The academies could be asked to evaluate the substantial evidence about the serious risks associated with Frontier AI that are in front of us

right now, such as bioweapons, cybersecurity, and loss of control over AI systems—all of which have been identified as serious concerns by Anthropic’s CEO, Dario Amodei. (They could also be asked to tackle the difficult-to-define risks that are likely to appear with the expected development of superintelligent machines.) They would use the well-established methods of quantitative risk assessment (also known as probabilistic safety analysis) to offer us a clear idea about the likelihood and consequences of such risks, as well as the level of uncertainty about the assessment and the degree of confidence we can have in the results. Finally, the academies are able to produce such a report in a relatively short period of time.

This solution worked for Hans Bethe many years ago. It can work for us again now.