Google DeepMind hires nuclear experts to assess AI risks

The expertise DeepMind wants

The minimum bar is a PhD in a field such as plasma physics or nuclear engineering, or comparable practical experience. Candidates also need experience working with a US government laboratory and knowledge of policies related to maintaining the nuclear arsenal, nonproliferation, arms control or threat reduction.

Preferred qualifications include AI safety experience, computational modeling, or work at large laser facilities. The role also calls for developing risk-assessment and mitigation methods that avoid classified material and export-controlled information. The posting suggests using dual-use scientific problems as substitutes for weapons-related topics.

The salary range is approximately $174,000 to $252,000, plus a 15% bonus.

Testing the boundary

The specialist would assess whether advanced AI models could help users with topics including fission and fusion, thermonuclear weapons, radiation hydrodynamics and magnetohydrodynamics. The evaluations must stay clear of classified information and export controls.

That scope makes the job more than a conventional safety review. DeepMind is recruiting people who can judge whether a model’s output has practical value in a highly sensitive field—and design tests that probe the risk without reproducing restricted material.

I think the posting signals that Google considers nuclear-related model assistance important enough to build dedicated expertise around. It does not establish that its tools can provide actionable instructions for making nuclear devices; the announcement leaves that threshold unclear. What it does show is a willingness to spend significantly on assessing the possibility.

The hard question is how to make a test realistic enough to reveal dangerous capability while keeping the test itself within safe boundaries. The job description names that constraint, but not how DeepMind plans to resolve it.