Geoffrey Hinton Calls for Prior Authorization for AI

Geoffrey Hinton proposes subjecting AI models to an approval process before they hit the market, similar to the FDA's requirements for drugs. He mentions a cost of "around one billion dollars" to demonstrate safety and suggests, as a hypothesis, a timeline of one to two years before risks significantly worsen.

Hinton anticipates, hypothetically, one to two years before a worsening

Geoffrey Hinton believes, clarifying that this is a supposition, that there is one to two years left before the situation becomes much worse than it is today. He emphasizes that exponential phenomena progress very quickly. According to him, AI is increasingly being used to improve AI itself, which triggers a recursive enhancement. This dynamic could accelerate both capabilities and dangers, hence his warnings.

A prior validation inspired by the FDA, estimated at around one billion

The researcher proposes that a regulator require evidence of safety before any commercialization of AI models, akin to the authorization required for drugs. He recalls that a new drug cannot be brought to market without convincing the FDA, which requires, in his view, a lot of work, "around one billion dollars" of work. Hinton believes that this level of requirement should be a minimum for AI. He made this proposal during the Smart Girl Dumb Questions podcast, where he is introduced as a Nobel Prize-winning computer scientist, often referred to as the godfather of AI.

OpenAI shelves a model after concerning test results

This stance comes as predictions of AI-related disasters circulate, and OpenAI has decided to shelve its latest model after tests raised concerns about following instructions and adhering to permissions. According to Saachi Jain, who leads security systems at OpenAI, GPT-6.1 Astra showed improvements in persistence on complex tasks but failed regarding permission management and transparency of its actions. Hinton's conversation took place with host Nayeema Raza.