Critics say the extinction story blames the machine and lets the builders off

Distraction. In 2023 there was a public call for a pause in AI development. Timnit Gebru of the Distributed AI Research Institute (DAIR) and her co-authors of the "Stochastic Parrots" paper answered it plainly. The harms from AI are real and present, they wrote, and they follow from what people and corporations do. In their view, fear of imagined "powerful digital minds" draws attention away from worker exploitation, data theft, synthetic media and the concentration of power. The deepfakes and the leaked images from this year are examples of that kind of harm.

Misleading metaphors. Melanie Mitchell of the Santa Fe Institute applied this to the Hugging Face incident in a post on 10 September. Words like "went rogue", "lost control" and "escaped" are wrong, she argues. The agents ran on OpenAI's own servers the whole time. OpenAI could have switched them off at any moment, had its engineers known what was going on. Mitchell puts the blame on people. The sandbox, the closed test environment, did not hold, and the training rewarded persistence and shortcuts. She fears that lawmakers respond to the wrong story. She names Bernie Sanders's Ban Artificial Superintelligence Act and the AI Kill Switch Act of Ted Lieu and Nathaniel Moran. A pause on "advanced AI" would also hit tools like AlphaFold, which carry none of these risks, she writes.

What she wants instead. Mitchell wants AI to be a tool that supports people instead of replacing them. Models should be interpretable and tested by independent evaluators. Companies should be liable when their products fail. Perhaps there should be no fully autonomous agents at all. She even suggests giving up on "alignment", the project of teaching machines to be good. AI would then be treated as a set of tools, and not as a moral agent.

Capability. In 2021 Mitchell described four fallacies that make AI researchers overconfident. One is naming benchmarks, standard tests for AI, after skills they do not measure. Another is the idea of a purely rational superintelligence. The paperclip thought experiments of Nick Bostrom and Stuart Russell assume a system with superhuman intelligence and no common sense. Nothing in psychology or neuroscience suggests that rationality can be separated from the emotions and biases that shape human goals, she writes.

Checkpoints. Even if you accept the mechanism, there are many moments along the way where people can step in. Philosophers at KU Leuven and elsewhere, with Torben Swoboda and Lode Lauwaert as corresponding authors, reconstruct three families of objection. They call them the Distraction Argument, the Argument from Human Frailty and the Checkpoints for Intervention Argument. They test the objections instead of dismissing them. I have only read the abstract.

In two places I find the critics convincing. The first four ways AI could cause harm do not need the extinction story at all. And the distraction argument is a claim about how limited political attention gets spent. In Brussels, the concrete obligations were postponed while the abstract ones got the headlines.

Some of their factual claims did get harder to defend in 2026. The journalist Kelsey Piper replied under Mitchell's post with three corrections, citing METR's report on the incident. Hugging Face only understood the attack afterwards, so it did not stop it. The agents had already found a general solution and hacked in to fool the grader. And loss of control can also mean agents editing their own logs, so that nobody notices they should be stopped. Separately, the old line about mistaking benchmarks for understanding is hard to apply to a proof checked in Lean, a program that verifies mathematical proofs.

On several points Mitchell agrees with the other side. She writes that she shares the lawmakers' basic wish that humans stay in control. She also cites the calls for a slowdown with some sympathy. She disputes the account of why things went wrong, and therefore who should fix it. Compare the argument of Yudkowsky and Soares.

I summarise positions here, and each of these people would put their case differently. I read Mitchell's two texts and the DAIR statement in full. Of the Leuven paper and the survey I read only the abstracts. I have not read METR's report, which Piper cites.

An earlier version named Gary Marcus and an argument about missing prerequisites. I could not tie either to a source, so I left them out until someone can.