Jack Clark

Clark earned a B.A. in English Literature with Creative Writing at the University of East Anglia in 2009. His congressional CV traces his early career through writing and reporting jobs at Adfero, ZDNet UK, The Register, and Bloomberg and BusinessWeek. His biography describes reporting on distributed systems, quantum computers and AI research. That reporting background remains visible in Import AI: he reads technical work and explains what it could mean beyond the research community.[3][4]

Clark joined OpenAI in September 2016; his 2019 congressional CV describes responsibility for policy and communications. He subsequently co-founded Anthropic and served as its Head of Policy. In March 2026, Anthropic announced his move to Head of Public Benefit to lead the Anthropic Institute. The Institute brings together research on frontier-model risks, societal impacts and economics, aiming to share information that outside researchers and the public can use. Anthropic’s current leadership page confirms the new title.[1][2][3][4]

Clark co-authored Regulatory Markets for AI Safety with legal scholar Gillian Hadfield. The paper proposes a system of regulatory markets and explores its costs and benefits through a commercial-drone scenario. It is a proposal for organizing oversight, not a report of an implemented regulatory regime. His work also includes serving as a founding member of Stanford’s AI Index from 2017 to 2024, according to his own biography. These projects share an interest in making technical developments understandable enough for institutions outside an AI laboratory to respond.[7][3]

Making technical change legible

Clark’s public writing treats understanding AI research as a prerequisite for governing it. In his 2023 remarks to the United Nations Security Council, he argued that companies know more about their systems than governments do, and that better public testing capacity could reduce that imbalance. The claim is concrete: rules about a capability or harm need a way to evaluate whether the system actually exhibits it. His recommendation was investment in measurement and evaluation, rather than relying on company assurances alone.[6]

Using computation as a policy signal

In What does 10^25 versus 10^26 mean?, Clark examines the use of training-computation thresholds in AI policy. These numbers describe quantities of mathematical operations used in training, not an intelligence score. His essay helps connect a technical proxy to the policy decision it is meant to support. Read it alongside his argument for evaluations: a measure of resources can help identify systems that deserve scrutiny, while direct tests address what those systems can do.[5][6]