Google AI chief Hassabis: Advanced AI that doesn’t pass review cannot be launched in the US

ME AI message, according to Beating monitoring, Google DeepMind CEO Demis Hassabis said that AGI may be only a few years away. He argued that the United States should take existing voluntary model testing and gradually turn it into a mandatory market-access regime. Under his plan, leading AI companies should initially submit their models for review at most 30 days before model release. After the assessment mechanism matures, models that meet the capability threshold must pass testing before they can be deployed in the U.S. market.

He suggested setting up a standards body similar to the U.S. financial industry regulator FINRA. The institution would be overseen by the federal government, with funding mainly sourced from AI companies, and it would include independent experts as well as representatives from open-source communities.

The testing focus would include network attacks, biological threats, deceptive behavior, and attempts to bypass safety restrictions. The standards would be updated regularly, and would gradually adopt independent tests that laboratories cannot see in advance, to prevent models from gaming the scores on public benchmarks.

The rules would be drawn based on model capability, without distinguishing between U.S. or foreign companies, and without distinguishing between open-source and closed-source. Startups and academic projects that do not meet the frontier threshold could be exempted.

The U.S. currently already allows frontier companies to voluntarily submit models to government testing up to 30 days before release. But an executive order issued in June explicitly excludes mandatory licensing and pre-approval. Hassabis argued for transforming this voluntary mechanism into a mandatory “pass,” and when risks are serious, the U.S. could coordinate with frontier labs to collectively slow down.

(Source: BlockBeats)