Companies do not have to wait for a catastrophe before they take AI risk seriously. Geoffrey Hinton thinks a serious incident may be what finally forces the labs and their regulators to move. He also says he is not a doomer, that his gut is we are still in control, and that a big effort might well work. If you are the one deploying the technology, take that second half and use it: choose tools and controls that still function when the model is gone.
Fate is a poor operating plan. Effort is not.
That warning comes from TechRadar’s report of a Guardian interview dated November 3, 2026. Hinton, a Nobel prize-winning computer scientist and one of the godfathers of AI, compared the kind of shock he has in mind to Three Mile Island, the 1979 partial meltdown that made a nuclear industry feel a danger it had already been warned about. The analogy is about how a crisis teaches. It is not a prediction that AI melts down a reactor. The incidents he described were an attack that shuts down global financial systems, or one that knocks out power generation and leaves massive brownouts. In TechRadar’s telling, regulation could only be a catastrophe away.
They understand the risks, he says. They are racing anyway.
Hinton is blunt about why caution is not the default among the people at the front of the race. TechRadar quotes him:
I think Zuckerberg and Musk and Sam Altman, they understand the risks, and they don’t give a shit. They’d rather have the money. That’s my belief.
That is his belief about three named people. It is not a strategy memo, and it is not a company list he issued. Mark Zuckerberg, Elon Musk, Sam Altman. You can infer the organizations. The sentence does not.
If the risks go unmanaged, the harms TechRadar reports him fearing are social as well as technical: “a huge gap between rich and poor, huge cultural problems, massive unemployment, just society being torn apart.” The pace still surprises him. In earlier comments cited by TechRadar, he said he didn’t think we would get here in only 40 years, and that 10 years ago he didn’t believe we would get here at all.
Buyers are not required to share the surprise, or the hurry. You cannot vote those three men into caution. You can refuse to build an operation that only works while the model is up.
What still works when the model is gone
The evidence for acting before a crisis is not a grid scenario. It is what happens to skill when a general-purpose chatbot does the thinking, and then leaves.
In a November 2026 Stanford Report, Lily Fesler, a senior research associate at the AI Hub for Education within the university’s SCALE Initiative, and colleagues went looking for studies rather than demos. Her team reviewed more than 800 papers on AI in education and focused on 20 high-quality causal studies. The account says the research is limited. Within that limit, the exam finding is uncomfortable for anyone treating a chatbot rollout as free speed.
The review found that the most detrimental tools were general-purpose chatbots. When students used one to study for an exam, it had the potential to decrease their test scores once they no longer had access.
A separate Stanford study, by Guilherme Lichand, an assistant professor at the Graduate School of Education, and predoctoral fellow Sachin Allums, looked at Brazilian middle schoolers split into four groups with different AI access. Two of the tasks were creative. Students who used an AI tool on the first task and then lost it did not keep the advantage. Their creative performance fell below students who had never had the tool. The damage was not evenly spread.
These effects are concentrated in the kids who use the AI tool more in the first task, so it suggests that the more you offload your abilities, the bigger the negative effects later on.
Lichand’s plainer line is the one that belongs on a rollout plan: “If I can completely delegate a task to AI and I never spend time thinking about it, then there’s no learning.” The pattern has a name, cognitive offloading. Shift the thinking onto the tool, and the tool takes the practice with it.
These are separate results. Neither one is a number you can drop on a board slide. The exam finding comes from a limited review of causal studies, and it reports no effect size. “Had the potential” is the claim. The creative finding is a different experiment, with middle schoolers, not a test of your staff. Do not merge them into a single forecast about productivity.
Design is the counterpoint. You do not have to swear the tools off. Fesler has said AI “has the potential to both improve and harm learning outcomes, ” and that “the difference likely comes down to the type of tool that’s being used.” Tools that ask Socratic questions, or that give feedback on a piece of writing without writing it, did not show the detrimental offloading effects as clearly. Hariharan Subramonyam, also an assistant professor at Stanford’s Graduate School of Education, puts the principle in four words: “Learning is representing.” Giving learners the capacity to create things, he argues, is where AI can make learning more meaningful and intentional.
The studies are about students. The buying distinction still travels. A chatbot that finishes the task and a tool that makes a person draft, then critiques the draft, are not the same purchase, even if both invoices say AI. The first will win the pilot because it feels done. The second is more likely to leave a skill behind when access ends. A pilot that only clocks speed while the tool is on will miss what these studies actually point to. The loss shows up after the tool is gone.
None of this requires a meltdown before the discussion is on the calendar. Treat a possible attack on financial systems or power as a resilience problem: vendor concentration, incident response, and a human on anything that spends company money. And ask the question the classroom evidence supports, even with its limits. If this tool disappeared tomorrow, who could still do the job?
Questions worth asking before the next rollout
-
Is Hinton saying nothing short of a catastrophe will make anyone take AI risk seriously?
Not in the account that holds up. TechRadar’s report of his Guardian interview says a serious incident may be what finally pushes companies and regulators to respond. He has also said he is not a doomer, that his gut is we are still in control, and that a big effort might well work.
-
What kind of incident is he actually describing?
An AI-enabled attack on financial systems or power generation, set beside the way the 1979 partial meltdown at Three Mile Island made a known danger feel real. The nuclear accident is his comparison, not a predicted AI meltdown.
-
Who does he think will keep racing anyway?
His stated belief, as TechRadar quotes it, is that Zuckerberg, Musk, and Sam Altman understand the risks and would rather have the money. He named three people. That is not a finding about a corporate plan.
-
Do general-purpose chatbots leave people better at the work once the tool is gone?
Stanford’s review of 20 causal studies, drawn from more than 800 papers, found general-purpose chatbots the most detrimental tools examined. Using one to study had the potential to lower later test scores after access ended. No size of drop is reported, and the research is limited.
-
Is every AI learning tool a problem?
No. Fesler has said the same technology can improve or harm learning depending on the tool. Assistants that ask questions, or that critique work without doing it, did not show the same offloading harm as clearly.
-
If this tool disappeared tomorrow, who on the team could still do the job?
That is the check worth running before you buy. In the Stanford creative study, students who leaned on AI and then lost it performed below students who never had it, and the harm was larger where the offloading had been heaviest. Those were middle schoolers, not your staff. The question still belongs on the rollout.
A big effort, in Hinton’s phrase, might well work. Inside a firm it looks unglamorous: which tool you buy, a human check on anything that spends, and the slightly annoying decision to make people do some of the reps. Three Mile Island taught a generation that a near-miss can change an industry. It also taught everyone else that you would rather not be the plant where the lesson lands.