Demis Hassabis Proposes a U.S. Body to Test Frontier AI Systems Before Release

Demis Hassabis Proposes a U.S. Body to Test Frontier AI Systems Before Release

The head of Google DeepMind has proposed creating a U.S. organization that would define the threshold for frontier AI and test models that cross it before release. Testing would initially be voluntary. If the methods prove effective and reliable, Hassabis proposes making approval a condition for deploying a model in the U.S. market.

On July 14, Hassabis published an essay on rules for frontier AI systems. Under his framework, a model would qualify as “frontier” if it achieved a specified score on evaluation tasks. The new organization would set this score and regularly update the tasks as model capabilities improve. This process would determine which laboratories become subject to the additional testing regime.

Hassabis uses FINRA as a model. FINRA is a private U.S. organization that oversees brokerage firms under the supervision of the U.S. Securities and Exchange Commission. He proposes that the new body’s board include independent technical experts as well as representatives from government, industry, and the open-source software community.

“Initially, laboratories would voluntarily submit models to the body for testing no more than 30 days before release.”

The evaluations would assess cyber risks, biological threats, and other high-risk areas. Hassabis also identifies nuclear risks as a possible threat. Separate tests could look for attempts to bypass built-in safeguards or for signs of deception. If the testing methods prove effective and reliable, Hassabis proposes making approval a condition for releasing a frontier model in the U.S.

CAISI, a center within the U.S. National Institute of Standards and Technology, has already received access to Google DeepMind models before their public release to evaluate biological risks, cyber threats, and risks to critical infrastructure. Under Hassabis’s proposal, the new body would set the threshold that determines whether a model must undergo testing.

If testing becomes mandatory, the body would define the threshold for a “frontier” model and the set of evaluations it must pass before entering the U.S. market. Hassabis also allows for coordinated slowdowns in development among laboratories working on frontier-class models if circumstances require one.

Large-scale evaluations require computing resources and specialists, so Hassabis expects industry to provide funding. The initial testing methods would be developed in consultation with frontier AI laboratories. The body would then build the capacity to create tests independently of the laboratories and keep those tests confidential from developers. This would reduce the chance that a model could be optimized for known questions. The future body’s independence will depend on its budget, the composition of its board, and its ability to develop its own evaluations.

Hassabis defines artificial general intelligence as a system with the full range of human cognitive abilities and expects it to accelerate science, medicine, and drug discovery. His proposal would require frontier models to be tested before release in the U.S. market.