Expert consensus on mitigations for the most severe AI risks
We are looking for research fellows to run expert consensus (Delphi) studies on the mitigations for the AI risks we are most worried about:
sudden loss of control during recursive self-improvement;
gradual loss of control as decisions are delegated to AI across the economy and the security state;
a small group seizing decisive advantage, for example through an AI-enabled coup; and
misuse of AI by rogue actors to cause mass harm.
For each of these risks, we want to know which mitigations the relevant experts rate as both effective and feasible, so that policymakers and companies have a clear, credible shortlist to act on.

You would take one risk and deliver the first wave of its Delphi study: assembling the expert panel, adapting our survey materials to the risk, running the first survey round either via Qualtrics or interview, and analysing initial themes. Presuming we could not complete all three rounds within 10 weeks, the fellow would write the materials for waves two and three, which our team would deliver, with the fellow continuing as an author on the resulting paper. The method, survey infrastructure, analysis pipeline, and expert networks already exist from our previous Delphi studies, so the fellow's time goes into execution and writing rather than design from scratch.

This would suit someone with working knowledge of at least one of the four risks.

We would expect the person to have completed BlueDot Impact’s Frontier AI Governance course, or equivalent. We also expect them to have demonstrated strong agency and decision-making running independent projects in the past. Strong, clear writing is also expected.