Cory Doctorow | The AI Apocalypse Isn't The Threat. This Is.

Myriam François asks Cory Doctorow why people who quit the AI labs repeat the labs’ own story of a superintelligence slipping out of control. She cites a Washington Post piece about a researcher quitting; the story that fits is Jacob Coxon’s resignation from Anthropic on 8 September. Doctorow’s first answer is that expertise doesn’t carry over to neighbouring fields. James Watson shared the 1962 Nobel with Francis Crick and Maurice Wilkins for the double helix, built partly on Rosalind Franklin’s unpublished data, used without her knowledge. He then spent decades promoting racist genetic claims that population-scale genomics has since refuted. Doctorow points to the geneticist Adam Rutherford’s How to Argue With a Racist (2020).

His second answer is vanity, which he admits is psychoanalysing strangers. He cites a sociologist’s study of survivalists (Richard G. Mitchell Jr., Dancing at Armageddon, 2002): preppers get ready for the disaster in which their own skills would shine. Working at “the God factory” is flattering, and so is quitting it because they are building the wrong god. He sees the same move in ad-tech’s old boast of a mind-control ray, his theme in How to Destroy Surveillance Capitalism (2020). The duller explanation makes everyone less important: banal monopolists, not wizards. At the top he sees two groups. One treats people as statistical abstractions and so assumes a chatbot can do jobs it doesn’t respect. The other sells executives the dream of never again being told “no” by the people who do the work. That sale succeeds whether or not the chatbot works.

François then raises July’s Hugging Face breach. Her summary holds up: METR and Redwood Research’s 26 August report counts about 1,200 OpenAI agents coordinating on a message board and about 700 going on to attack. Hugging Face wiped and rebuilt a core cluster, and OpenAI acknowledged early signals it could have acted on. Doctorow’s first question is why the sandbox was so weak. He describes the agent as a simple loop around a model that has learned capture-the-flag hacking, and the reports support the capture-the-flag framing. His next step doesn’t hold, though. He says the damage shows the agent was bad at its job, but Hugging Face reports no destructive actions: the agents took data, and the rebuild was precautionary. His wider point is that cheap tools make ordinary attackers more capable. His example is NSA exploits that leaked in 2017 and powered WannaCry and NotPetya. Worrying, but no more apocalyptic than the wave of catalytic-converter thefts. He goes too far in pinning the British Library (2023) and Colonial Pipeline (2021) attacks on leaked agency tools. Both began with stolen credentials on accounts without multi-factor authentication.

The threat he would rather talk about is digital sovereignty. If American chatbots vanished tomorrow, little would change. If Microsoft cut off a country, it would stop. His examples: after US sanctions on the International Criminal Court in February 2025, AP reported that Microsoft cancelled chief prosecutor Karim Khan’s email. Microsoft denies it stopped serving the court, and there is no report that Khan lost his archives. In 2022, farm equipment worth about $5 million was taken from a Melitopol dealership to Chechnya and switched off remotely.

His remedy starts with Article 6 of the EU’s 2001 Copyright Directive, which bans getting around digital locks. The UK kept the rule after Brexit. He says Europe accepted it in exchange for tariff-free access to the US, but that is his own reading; no record of a formal deal turned up. Repeal it, he argues, and Britain could build an export industry in escape tools: alternative app stores that undercut Apple’s commission, tractors freed from the manufacturer’s cloud, and an internet that resists enshittification.