Geoffrey Hinton calls for mandatory safety reviews before AI products are launched.

Nobel laureate and computer scientist Geoffrey Hinton said on the "Smart Girl Dumb Questions" podcast on Tuesday that AI companies should pass safety reviews by regulatory agencies before releasing new products. He compared AI regulation to drug approval, noting that pharmaceutical companies cannot bring new drugs to market without permission, but instead need to conduct extensive work and convince the U.S. Food and Drug Administration of their safety; AI should at least meet similar requirements. Hinton made these remarks amid intense debate in the AI industry over model safety. Previously, OpenAI decided to postpone the release of its GPT-6.1 Astra model after testing found it failed to meet requirements in permission control, scope of action, and explanation of results. OpenAI's head of safety systems, Saachi Jain, said that although Astra is better at continuously completing difficult tasks, it still falls short in adhering to authorizations and explaining its own behavior. Hinton also warned that AI is being used to improve AI itself, potentially leading to recursive self-improvement that accelerates both capabilities and risks at the same time. He emphasized that this is only his personal judgment, but he expects the situation could visibly worsen in the next year or two, so regulatory and safety verification mechanisms should be established as soon as possible.