Research Scientist, Manipulation Evals

Role overview Shape and deliver research evaluations that help assess when AI systems can manipulate people in harmful ways. You will set priorities, develop the evaluation methods and evidence, and publish findings that can inform regulatory decisions. The work combines independent research with engagement from policy and technical stakeholders.

Responsibilities - Identify manipulation risks that are important and practical to evaluate, and define what harmful manipulation means for the work. - Build evaluations, including behavioral tests, measurement approaches, and benchmarks, to assess model capabilities and risks. - Run experiments, analyze results, and turn findings into rigorous evidence and research publications. - Work with external researchers, policy stakeholders, and partner organizations to ensure methods address real needs. - Present and defend conclusions, including their limitations, to technical and nontechnical audiences.

Requirements - Strong research experience in AI safety, machine learning, behavioral science, or a closely related area. - A record of producing rigorous research, such as peer-reviewed work or substantial independent projects. - Ability to design evaluations and analyze evidence about complex model behavior. - Sound judgment in selecting research questions and communicating uncertainty. - Clear written and verbal communication, with the ability to collaborate across disciplines.

Benefits and work setup - Full-time, remote role; Europe and San Francisco are preferred locations. - Europe-based compensation is listed at $130,000–$200,000 gross annually, depending on experience; San Francisco pay is adjusted for local costs.