Gates urges US to mandate AI safety reviews

Bill Gates argues that the US government must move beyond self-regulation to mandate safety reviews for new AI models, noting that the US currently lacks mandatory review while other governments, including China, are waiting for US leadership. He states that the current administration has shown little proclivity for regulation and favors self-regulation, creating a stasis that leaves society vulnerable. Gates highlights that while closed models have built-in safeguards, open-weights models allow attackers to remove monitoring and safeguards, making them dangerous for cyberattacks. He notes that the Hugging Face hack by OpenAI highlighted the urgency of control issues, but he believes the near-term challenge is malicious actors using unsafeguarded models.

Gates advocates for third-party auditing, such as by the UK Safety Institute, to verify safeguard strength and prevent jailbreaking. He emphasizes that without government involvement, it is difficult to ensure that control problems are solved before next year's major model releases. Gates argues that self-regulation is insufficient due to the number of competitors and the severity of risks, and that the US must lead on mandatory safety standards to prevent catastrophic outcomes.