Geoffrey Hinton Wants AI Models to Face FDA-Style Safety Approval Before Release

October 7 (IT Home) — Geoffrey Hinton wants AI companies to clear a regulatory hurdle that the pharmaceutical industry has long been familiar with.

On the podcast Smart Girl Dumb Questions, released on Tuesday local time, the computer scientist and Nobel laureate known as the “godfather of AI” proposed that AI companies should first have to prove to regulators that their products are safe before releasing them.

“You can’t just develop a new drug and put it on the market,” Hinton told podcast host Nayeema Raza. “You have to convince the Food and Drug Administration. And to do that, you need to do a huge amount of work, and it costs around a billion dollars.”

“At the very least, there should be that kind of requirement for AI as well,” Hinton added.

Hinton’s proposed solution to the AI safety debate comes against a backdrop of mounting predictions of catastrophic outcomes from AI — and of OpenAI shelving its latest model after testing, out of concern about whether the model would obey user instructions and respect permission boundaries.

IT Home notes that Saachi Jain, head of OpenAI’s safety systems, told Business Insider earlier that the GPT-6.1 Astra model had improved at sustaining work on difficult tasks, but still had shortcomings in permission control and behavioural logging feedback.

Hinton argues that AI is increasingly able to iterate and improve itself, which could both accelerate gains in capability and amplify potential risks.

“Exponential growth is astonishing,” Hinton said. “We’re already starting to see recursive self-improvement: people are using AI itself to improve AI in all sorts of ways.”

“Of course, this is just my guess,” Hinton added. “But I reckon it’s only about one or two years before the situation is far worse than it is now.”