Stuart Russell

Stuart Russell is an AI researcher at the University of California, Berkeley. His work represented here spans how agents infer and optimize rewards, how they can remain responsive to human guidance, and a research agenda for beneficial AI. He has also taken public positions on autonomous weapons and advanced-AI risks.^[3][8][11]

His technical work represented here includes Algorithms for Inverse Reinforcement Learning with Andrew Ng and Policy invariance under reward transformations: Theory and application to reward shaping with Ng and Daishi Harada.^[5][6] The first paper frames the problem of inferring a reward function from observed behavior and shows why that inference is underdetermined: several rewards can fit the same behavior. The second gives a condition for adding a shaping reward without changing which policies are optimal. Neither result identifies which human objective ought to be optimized.^[5][6]

Beneficial AI as a research agenda

With Daniel Dewey and Max Tegmark, Russell mapped research problems across verification, validity, security and control, alongside economic and legal effects, in Research Priorities for Robust and Beneficial Artificial Intelligence. It was a proposed agenda, rather than a demonstration that those problems had been solved.^[3] In a 2016 Science Friday interview, he argued that choosing the right objectives was itself part of AI research and discussed Berkeley’s Center for Human Compatible Artificial Intelligence.^[7]

Learning with irreversible consequences

Russell also coauthored Safe Learning Under Irreversible Dynamics via Asking for Help with Benjamin Plaut, Juan Liévano-Karim and Hanlin Zhu. The JMLR 2026 paper studies agents that learn without resets and can ask a mentor before acting. Its theoretical objective accounts for mistakes made while learning, rather than only the eventual policy. The guarantees are relative to a mentor and depend on the paper’s model assumptions; the authors say the algorithm is not ready for practical use.^[2]

Autonomous-weapons advocacy

FLI’s July 29, 2015 report identifies Russell and Toby Walsh as announcing the 2015 autonomous-weapons open letter at an IJCAI press conference in Buenos Aires. This documents his role in that appeal for restrictions on military autonomy, beyond his technical work on reward learning. The open letter advocated a ban on offensive autonomous weapons beyond meaningful human control; the announcement documents Russell’s participation, without measuring the campaign’s effect.^[8]

Cooperative assistance and correction

With Dylan Hadfield-Menell, Anca Dragan and Pieter Abbeel, Russell coauthored Cooperative Inverse Reinforcement Learning and The Off-Switch Game. The first models teaching and learning as joint pursuit of a human reward; the second studies when uncertainty and informative human choices make oversight valuable. These formal contributions connect his objective-selection research aim to specific incentive models, while leaving their human-model and computation assumptions in place.^[9][10]

Russell also coauthored Should Robots be Obedient?, which studies how errors in an agent’s model of human preferences affect the tradeoff between obeying and acting autonomously. These are formal and simulated models, not evidence that deployed agents reliably accept correction.^[1]

Public risk appeals in 2023

Russell appears in the selected signatories of the 2023 pause letter on giant AI experiments and in the original announcement of the 2023 statement on AI risk. The first proposes a training pause; the second prioritizes extinction-risk mitigation without prescribing that intervention.^[11][13]

In an April 2023 Science Friday interview, he distinguished concerns about stronger future systems from his assessment of current systems. His risk argument was a forecast, rather than a claim that contemporary systems had already caused that outcome.^[14]

Selected authored research

• Policy invariance under reward transformations: Theory and application to reward shaping (with Andrew Ng and Daishi Harada): conditions for policy-preserving reward shaping.^[6]
• Algorithms for Inverse Reinforcement Learning (with Andrew Ng): the inverse problem of inferring rewards from behavior.^[5]
• Research Priorities for Robust and Beneficial Artificial Intelligence (with Daniel Dewey and Max Tegmark): a multidisciplinary safety research agenda.^[3]
• Cooperative Inverse Reinforcement Learning and The Off-Switch Game (with Dylan Hadfield-Menell, Anca Dragan and Pieter Abbeel): formal models of assistance and human intervention.^[10][9]
• Should Robots be Obedient? (with Smitha Milli, Dylan Hadfield-Menell and Anca Dragan): the effect of misspecified human models on autonomous action.^[1]
• Safe Learning Under Irreversible Dynamics via Asking for Help (with Benjamin Plaut, Juan Liévano-Karim and Hanlin Zhu): mentor-guided learning where early mistakes cannot be reset.^[2]