An Independent Body is Needed to Stop Out-of-Control AI: Google's Proposal

Demis Hassabis, CEO of Google DeepMind, published a lengthy document on Tuesday calling for the creation of an independent body tasked with evaluating the most advanced artificial intelligence models before they hit the market. The text, titled "A Framework for Frontier AI and the Dawning of a New Age," proposes a structure modeled after the Financial Industry Regulatory Authority (FINRA), the authority that oversees the financial sector in the United States.

The idea is to create a "standards body" that, in its initial phase, would receive models in preview up to 30 days before their launch to subject them to tests on cybersecurity, biological risk, and the ability to evade integrated security measures. Once the effectiveness of the evaluation protocol has been demonstrated, Hassabis writes, compliance could become mandatory for anyone wishing to distribute a frontier model in the United States, with the laboratories also involved in managing vulnerabilities discovered after release.

Hassabis wants a body to stop out-of-control AI. The proposal comes after ad hoc reviews conducted by the U.S. government on two models, Anthropic's Mythos and OpenAI's Sol, criticized for the technical incompetence of the evaluators and the opacity of the criteria used to decide whether to authorize their release. The new body envisioned by Hassabis would absorb that task, remaining under the umbrella of the federal government but with funding and management entrusted to the industry itself.

The political terrain remains slippery. Sriram Krishnan, a White House advisor on artificial intelligence and general partner at a16z, has already ruled out the creation of a dedicated federal regulator, comparing the hypothesis to a non-existent "FDA for AI." Building the standards body as a self-regulatory organization, modeled on FINRA, appears to be an attempt to circumvent that ideological resistance from the Trump administration toward new regulatory agencies.

In Hassabis's plan, the board should include independent technical experts, representatives from the open-source ecosystem, and delegates from funding laboratories. Benchmarks would be updated approximately quarterly, with a gradual transition to tests conducted by independent parties outside the laboratories themselves, to prevent models from being optimized solely to pass the expected tests.

Among the proposed measures are digital watermarking of images generated by AI and the request for more readable outputs for the interpretability of systems. The timelines remain less defined: Hassabis talks about AGI "probably a few years away" and describes the current phase as the "edges of the singularity," with an anticipated impact equal to "ten times the industrial revolution, ten times faster." Suggestive formulas, useful to impart urgency to the message, but which remain forecasts rather than verifiable certainties.

The CEO of DeepMind hopes that the initiative, although conceived to start in the United States, will encourage other countries to adopt shared standards, given that the effects of frontier artificial intelligence will not stop at American borders.