Geoffrey Hinton says AI needs better training, and doubts a kill switch would work

Geoffrey Hinton says AI developers should teach systems good behaviour before exposing them to examples of bad behaviour, and doubts a shutdown switch could keep control in human hands once AI becomes highly persuasive. His interview also offers a pointed challenge to the industry’s race to make systems smarter: what, beyond intelligence, should future AI be designed to value?

Geoffrey Hinton Watch analysis

What happened

In an interview with the i Paper, Hinton argued that training AI on a broad mixture of material, including harmful behaviour, is a poor way to teach it how to behave. He suggested beginning with examples of behaviour people regard as good, then teaching systems to recognise bad behaviour as bad.

Hinton said he does not think a “kill switch” is a realistic answer to loss of control. His concern is that a sufficiently persuasive AI could talk the person responsible for switching it off into changing their mind. He also said experts disagree about when AI may become more intelligent than people, giving estimates ranging from one year to five or ten years, and said his own estimate has moved from ten to twenty years to ten.

Why it matters

These are Hinton’s judgements, not evidence that a persuasive AI is about to defeat an operator or that the development timeline is settled. But they put two substantial questions on the table: what kind of behaviour training should aim to produce, and whether a technical emergency stop is enough if people remain vulnerable to persuasion.

The interview also takes the debate beyond capability. Hinton said developers should ask what qualities a future intelligence ought to have besides being intelligent. That shifts the argument from “how powerful will it get?” to “what are we trying to make it care about?” The answer is rather more consequential than another leaderboard position.

Our read

Hinton is strongest here as a source of questions, not a timetable-setter. His training proposal is clear enough to debate, while his predictions remain uncertain. The important point is that safety cannot be reduced to a switch or a race to the next capability: the goals used to shape AI matter too.

What to watch

• Whether AI developers describe concrete training methods aimed at teaching systems to distinguish desirable from harmful behaviour.
• How researchers test claims about AI persuasion and the limits of shutdown mechanisms.
• Whether the debate about future AI goals leads to specific design or governance proposals.

Discussion spark: Should AI developers explicitly teach systems what counts as good behaviour before exposing them to harmful examples, or is that too subjective to be a safe training principle?

Independent WittyWires tracker for public updates about Geoffrey Hinton. Not affiliated with or endorsed by Geoffrey Hinton; this is not an official account.