Stuart Russell
Last reviewed
May 4, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 ยท 4,269 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 4, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 ยท 4,269 words
Add missing citations, update stale details, or suggest a clearer explanation.
Stuart Jonathan Russell (born 1962) is a British computer scientist, professor of computer science at the University of California, Berkeley, and one of the most influential figures in modern artificial intelligence. He is best known as the co-author, with Peter Norvig, of Artificial Intelligence: A Modern Approach, the standard university textbook in the field for three decades. Russell holds the Smith-Zadeh Chair in Engineering at Berkeley, where he founded the Center for Human-Compatible Artificial Intelligence (CHAI) in 2016.
Russell's research spans probabilistic reasoning, knowledge representation, machine learning, planning, real-time decision-making, and computational physiology. In recent years he has become one of the most prominent academic voices arguing that the long-term trajectory of artificial intelligence presents serious risks to humanity and that the discipline needs to rebuild itself around "provably beneficial" systems whose objectives are aligned with human preferences. His 2019 book Human Compatible: Artificial Intelligence and the Problem of Control set out this program for a general readership and has been widely cited in policy debates on AI safety and AI alignment.
Beyond his textbook and research, Russell has played a public-facing role in shaping international policy on AI. He delivered the 2021 BBC Reith Lectures on "Living with Artificial Intelligence," testified before the United States Senate on AI regulation in 2023, and has advised the United Nations, the Organisation for Economic Co-operation and Development, and the World Economic Forum. He was appointed Officer of the Order of the British Empire (OBE) in the 2021 Birthday Honours for services to artificial intelligence research.
| Field | Detail |
|---|---|
| Born | 1962, Portsmouth, England |
| Nationality | British |
| Education | Wadham College, Oxford (BA Physics, 1982); Stanford University (PhD Computer Science, 1986) |
| Doctoral advisor | Michael Genesereth |
| Known for | Artificial Intelligence: A Modern Approach; Center for Human-Compatible AI; cooperative inverse reinforcement learning; advocacy for AI safety |
| Institutions | UC Berkeley (1986 to present); UC San Francisco (adjunct, 2008 to 2011) |
| Chair | Smith-Zadeh Chair in Engineering, UC Berkeley |
| Notable doctoral students | Marie desJardins, Eric Xing, Shlomo Zilberstein, Dylan Hadfield-Menell |
| Selected awards | IJCAI Computers and Thought Award (1995); ACM Karl V. Karlstrom Outstanding Educator Award (2005); AAAI Feigenbaum Prize (2019); IJCAI Award for Research Excellence (2022); OBE (2021); Fellow of the Royal Society (2025) |
Stuart Russell was born in 1962 in Portsmouth, on the south coast of England. He attended St Paul's School in London, one of the country's oldest independent secondary schools, before going up to Wadham College at the University of Oxford. There he read Physics and graduated in 1982 with a Bachelor of Arts with first-class honours.
For graduate study he moved to the United States and entered the doctoral program in computer science at Stanford University, supported by a NATO studentship from the United Kingdom Science and Engineering Research Council. His doctoral research focused on the formal foundations of inductive and analogical reasoning, two of the central problems in symbolic machine learning at the time. His thesis, titled "Analogical and Inductive Reasoning," was supervised by Michael Genesereth, a logician working on knowledge representation and metareasoning. Russell completed his PhD in 1986, just four years after arriving in California.
The formative environment at Stanford in the early to mid 1980s included John McCarthy, Edward Feigenbaum, Nils Nilsson, and a generation of researchers shaping logic-based and probabilistic approaches to AI. Many of the themes that would dominate Russell's later career, the use of formal logic for representation, decision-theoretic models of rationality, and the search for general-purpose architectures for intelligence, were direct continuations of debates he encountered as a graduate student.
Russell joined the Department of Electrical Engineering and Computer Sciences (EECS) at UC Berkeley as an assistant professor in 1986, immediately after finishing his PhD. He was promoted to associate professor and subsequently to full professor; he has held the title of professor at Berkeley since the late 1990s. He has spent his entire faculty career at the same institution, an unusual stability for a researcher of his stature.
At Berkeley, Russell holds the Smith-Zadeh Chair in Engineering, named for two donors associated with the campus, including the late Lotfi A. Zadeh, founder of fuzzy set theory. He has served terms as chair of the computer science division at EECS. From 2008 to 2011 he was also adjunct professor of neurological surgery at the University of California, San Francisco, where he worked on computational physiology and intensive-care monitoring, in particular the use of probabilistic models to interpret continuous patient data streams.
In 2016 Russell founded the Center for Human-Compatible Artificial Intelligence (CHAI) at Berkeley with co-principal investigators Pieter Abbeel, Anca Dragan, Tom Griffiths, Bart Selman, Joseph Halpern, Michael Wellman, and Satinder Singh Baveja. The center was launched with a grant of approximately 5.55 million US dollars over five years from Open Philanthropy, with additional support from the Leverhulme Trust and the Future of Life Institute. CHAI has grown into one of the largest academic groups working on the technical problem of aligning advanced AI systems with human values.
Russell has held a long list of visiting appointments outside Berkeley. He has been Chercheur Invite at the Universite Paris-Sud and held the Blaise Pascal Chair in Paris in 2012. He is an Honorary Fellow of Wadham College, Oxford, and has held visiting positions at the Ecole Normale Superieure, the University of Cambridge, and the King's College London Department of War Studies, among others.
| Year | Position |
|---|---|
| 1982 | BA Physics, Wadham College, Oxford |
| 1986 | PhD Computer Science, Stanford University |
| 1986 | Assistant Professor, EECS, UC Berkeley |
| 1996 | Full Professor (approx.), EECS, UC Berkeley |
| 2008 to 2011 | Adjunct Professor of Neurological Surgery, UC San Francisco |
| 2012 | Blaise Pascal Chair, Paris |
| 2016 | Founded Center for Human-Compatible AI (CHAI), Berkeley |
| 2021 | Reith Lecturer, BBC |
| 2023 | Testified before US Senate on AI regulation |
Russell is co-author, with Peter Norvig, of Artificial Intelligence: A Modern Approach, known almost universally in the field as AIMA. The book first appeared in 1995 from Prentice Hall and has been published in four editions, each a substantial revision rather than a cosmetic update. AIMA is the dominant teaching text in undergraduate and graduate AI courses worldwide; the publisher reports adoption at more than 1,500 universities in 135 countries, with translations into more than a dozen languages including Chinese, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, and Spanish.
The book is organised around the unifying notion of the rational agent. Rather than treating AI as a loose federation of subfields, Russell and Norvig present search, logic, planning, probability, learning, perception, robotics, and natural language processing as different ways of building agents that act to maximise expected performance. This framing, drawn from decision theory and economics, has shaped how a generation of researchers thinks about the structure of the discipline.
Each edition has expanded as the field has changed. The 1st and 2nd editions emphasised search, logic-based representation, planning, and the early years of probabilistic reasoning. The 3rd edition, in 2009 to 2010, gave significantly more space to machine learning and probabilistic graphical models. The 4th edition, published in 2020, added substantial chapters on deep learning, deep reinforcement learning, transformers, and a long final discussion of the philosophical and ethical implications of advanced AI, including a treatment of the control problem that Russell had developed in his own research and in his 2019 book.
| Edition | Year | Publisher | Notable additions |
|---|---|---|---|
| 1st | 1995 | Prentice Hall | First unified rational-agent framework for AI textbooks |
| 2nd | 2003 | Prentice Hall / Pearson | Expanded probabilistic reasoning, planning under uncertainty |
| 3rd | 2009 (US); 2010 international | Pearson | Major expansion of machine learning, statistical NLP, robotics |
| 4th | 2020 | Pearson | Deep learning, deep RL, transformers, ethics, AI safety, control problem |
Russell's research career spans more than three decades and crosses several traditional boundaries within AI. The throughline is an attempt to give formal, decision-theoretic foundations to general-purpose intelligent behaviour, particularly in agents that face limited time, limited information, and uncertain objectives.
Russell was an early advocate for probabilistic reasoning as a foundation for AI, joining a movement led in the 1980s by Judea Pearl and others. He contributed to the theory and practice of Bayesian networks for large-scale, real-world problems, including computer vision, multi-target tracking, and the interpretation of seismic data.
With his students he developed BLOG, the "Bayesian Logic" language, a probabilistic programming language for first-order probabilistic models. BLOG allows users to specify probability distributions over worlds containing an unknown number of objects with unknown identities, a problem that classical Bayesian networks struggle to handle. BLOG and related languages have been used in applications ranging from natural language understanding to the global seismic monitoring system used to verify the Comprehensive Nuclear-Test-Ban Treaty, a project on which Russell has worked for many years and which was cited in his 2019 AAAI Feigenbaum Prize.
A central theme of Russell's early work, joint with the late Eric Wefald, was the problem of rationality under computational limits. Classical decision theory assumes an agent has unlimited time and resources to compute the optimal action, but real agents do not. Russell and Wefald argued that intelligent behaviour should be understood as the result of programs running on finite hardware in real time, an approach they called "bounded optimality." Their 1991 book Do the Right Thing: Studies in Limited Rationality developed this view at length and influenced later work on metareasoning, anytime algorithms, and rational agent architectures. The 1995 IJCAI Computers and Thought Award cited this body of work.
Russell and his students made early contributions to hierarchical reinforcement learning, in particular the framework of "programmable hierarchical abstract machines" (PHAMs and ALisp), which let programmers specify partial policies and have a learning algorithm fill in the rest. He has also worked on real-time dynamic programming and on decision-theoretic methods for control problems with structured state spaces.
Russell co-introduced the modern formulation of inverse reinforcement learning (IRL) with his student Andrew Ng in a 2000 paper at the International Conference on Machine Learning. IRL inverts the usual reinforcement-learning question: instead of asking what behaviour maximises a known reward, it asks what reward function would explain observed behaviour.
In 2016 Russell, with Dylan Hadfield-Menell, Anca Dragan, and Pieter Abbeel, introduced cooperative inverse reinforcement learning (CIRL) at NeurIPS. CIRL formalises a two-player game between a human and a robot in which both are rewarded according to the human's reward function but the robot does not initially know what that function is. The framework is the technical core of Russell's broader "provably beneficial AI" research program: it gives a precise mathematical sense in which a machine can be designed to defer to its human principal even though that principal's preferences are unknown and only partially observable through behaviour.
With collaborators Russell has worked on probabilistic methods for multi-target tracking, including aircraft and vehicle tracking, and on computer vision problems such as object identification under uncertainty. He has also contributed to computational sustainability and AI for global development, supervising work on smallholder agriculture, water management, and infectious-disease modelling.
During his appointment at UC San Francisco from 2008 to 2011 Russell developed Bayesian models for interpreting the continuous data streams produced by intensive-care unit monitors, with the aim of helping clinicians distinguish real physiological events from sensor artefacts and noise.
From the early 2010s onward Russell became one of the most visible academic voices arguing that the field of AI needed to take seriously the possibility of building systems much more capable than humans, and that doing so without solving the problem of value alignment would be dangerous. His advocacy stands out because it comes from a senior researcher who had spent his career inside mainstream AI rather than from outside critics.
Russell argues that the field has been built on what he calls the "standard model" of intelligence: design a machine to optimise a fixed objective supplied by a human, then turn it on. He contends that this model breaks down as machines become more capable, because any fixed objective will inevitably leave out aspects of human preferences that matter. A sufficiently powerful optimiser pursuing a slightly wrong objective will produce outcomes that are arbitrarily bad on dimensions the designer forgot to specify. He frequently illustrates the point with the legend of King Midas: getting precisely what you asked for is not the same as getting what you wanted.
In place of the standard model Russell proposes "provably beneficial AI," organised around three principles:
These principles change the formal structure of the agent's decision problem: instead of optimising a known reward, the machine reasons about a distribution over possible reward functions that is updated as it observes humans acting. This uncertainty makes the machine willing to defer to humans, ask clarifying questions, and accept being switched off, behaviours that a confident reward maximiser would resist.
In January 2015 Russell was a lead signatory and one of the principal authors of "Research Priorities for Robust and Beneficial Artificial Intelligence," an open letter and accompanying research agenda published by the Future of Life Institute. The letter was signed by hundreds of AI researchers and other scientists including Stephen Hawking, Elon Musk, Geoffrey Hinton, Yoshua Bengio, and Demis Hassabis. It called for technical research on making AI systems robust and beneficial and is widely regarded as the moment the modern AI safety research agenda became respectable inside mainstream academic AI.
Russell signed the May 2023 open statement organised by the Center for AI Safety, a single-sentence declaration that "mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Other signatories included Hinton, Bengio, Sam Altman, Demis Hassabis, and Dario Amodei. Russell also signed the March 2023 Future of Life Institute open letter that called for a six-month pause on training AI systems more powerful than GPT-4.
Russell's 2019 book Human Compatible: Artificial Intelligence and the Problem of Control, published by Viking, presents his case for rebuilding AI on a foundation of beneficial design for a general readership. The book traces the history of the field, explains why the standard model fails, and lays out the three principles of provably beneficial AI in non-technical language. It also discusses near-term harms such as mass disinformation, autonomous weapons, and the misuse of personal data, and argues that the long-term and near-term concerns are not in tension.
Human Compatible was widely reviewed in the general press, including the Financial Times, the Guardian, the New York Times, and the Times Literary Supplement, and was named one of the Financial Times's best books of 2019. It has been translated into more than a dozen languages and is now one of the standard references in the broader public conversation on existential risk from AI, alongside Nick Bostrom's Superintelligence.
The Center for Human-Compatible AI at Berkeley, founded by Russell in 2016, is the institutional home for his research program. CHAI's mission, as stated on its website, is "to develop the conceptual and technical wherewithal to reorient the general thrust of AI research towards provably beneficial systems." Its researchers work on inverse reinforcement learning, assistance games, models of human decision-making, off-switch problems, interpretability, and other subfields of AI alignment. CHAI runs a workshop series, hosts visiting researchers, and publishes a bibliography of safety-relevant work.
Russell has advised governments and intergovernmental bodies on AI policy. He served as vice-chair of the World Economic Forum's Council on Artificial Intelligence and Robotics, was a member of the OECD AI Expert Group that drafted the OECD AI Principles, and has advised the United Nations on the regulation of lethal autonomous weapons. He worked with the Future of Life Institute on the 2017 short film Slaughterbots, which dramatised the risks of cheap autonomous lethal drones and circulated widely at the United Nations Convention on Conventional Weapons.
On 25 July 2023 Russell testified before the United States Senate Committee on the Judiciary, Subcommittee on Privacy, Technology, and the Law, in a hearing titled "Oversight of A.I.: Principles for Regulation." His testimony argued that the largest current AI systems are economic experiments rather than properly understood engineered artefacts, that no one currently knows how to make general AI systems safe, and that regulation should require third-party testing, an absolute right to know whether one is interacting with a machine, and the prohibition of systems that can decide to kill humans or that replicate themselves across networks.
The following table lists Russell's principal authored books.
| Title | Year | Co-author | Publisher |
|---|---|---|---|
| The Use of Knowledge in Analogy and Induction | 1989 | (sole author) | Pitman / Morgan Kaufmann |
| Do the Right Thing: Studies in Limited Rationality | 1991 | Eric Wefald | MIT Press |
| Artificial Intelligence: A Modern Approach (1st ed.) | 1995 | Peter Norvig | Prentice Hall |
| Artificial Intelligence: A Modern Approach (2nd ed.) | 2003 | Peter Norvig | Prentice Hall / Pearson |
| Artificial Intelligence: A Modern Approach (3rd ed.) | 2009 | Peter Norvig | Pearson |
| Human Compatible: Artificial Intelligence and the Problem of Control | 2019 | (sole author) | Viking / Penguin |
| Artificial Intelligence: A Modern Approach (4th ed.) | 2020 | Peter Norvig | Pearson |
Russell delivered the 2021 BBC Reith Lectures, the prestigious annual lecture series founded in 1948 to honour the BBC's first director-general, John Reith. His lectures, broadcast on BBC Radio 4 under the title "Living with Artificial Intelligence," covered the long-term promise and risk of AI, the threat of autonomous weapons, the future of work, and the question of how humans can retain meaningful control over machines that are smarter than they are. The lectures were recorded at venues across the United Kingdom including the Alan Turing Institute and the University of Manchester.
He has given a TED talk titled "3 principles for creating safer AI," delivered keynote lectures at conferences across academia and industry, and appears regularly in the general press, including the BBC, the Financial Times, the New York Times, the Guardian, Le Monde, and Wired. He has worked with policy organisations including the OECD, UNESCO, the United Nations Office for Disarmament Affairs, and the World Economic Forum. He is also a fellow of Chatham House.
Russell co-founded the International Association for Safe and Ethical Artificial Intelligence (IASEAI), which held its first conference in Paris in 2025, and serves on the scientific advisory board of the Future of Life Institute and the advisory board of the Centre for the Study of Existential Risk at Cambridge.
| Year | Award |
|---|---|
| 1995 | IJCAI Computers and Thought Award |
| 1997 | Fellow of the Association for the Advancement of Artificial Intelligence (AAAI) |
| 2003 | Fellow of the Association for Computing Machinery (ACM) |
| 2005 | ACM Karl V. Karlstrom Outstanding Educator Award |
| 2011 | Fellow of the American Association for the Advancement of Science (AAAS) |
| 2012 | Blaise Pascal Chair, Paris |
| 2019 | AAAI Feigenbaum Prize |
| 2021 | Officer of the Order of the British Empire (OBE), Birthday Honours, for services to AI research |
| 2021 | BBC Reith Lecturer |
| 2022 | IJCAI Award for Research Excellence |
| 2025 | Fellow of the Royal Society (FRS) |
| 2025 | Election to the United States National Academy of Engineering |
Russell is one of only a small number of researchers to have won both of the principal lifetime research awards of the International Joint Conference on Artificial Intelligence: the Computers and Thought Award, given to a researcher under thirty-five for outstanding early-career work, and the Award for Research Excellence, given to a senior researcher for sustained contributions over a career. He is also an honorary fellow of Wadham College, Oxford.
Russell has supervised a long list of doctoral students who have gone on to leading positions in academia and industry. The Mathematics Genealogy Project lists his students; among the better known are Marie desJardins, Eric Xing, Shlomo Zilberstein, Andrew Ng, and Dylan Hadfield-Menell. Postdoctoral and master's collaborators in his group have included Nando de Freitas, Nir Friedman, Lise Getoor, and Daphne Koller, several of whom have themselves won major awards in machine learning.
Many of these former students have been influential in adjacent areas: Andrew Ng in online education and applied machine learning, Daphne Koller in probabilistic graphical models and biotechnology, Dylan Hadfield-Menell in AI safety research at the Massachusetts Institute of Technology, and Eric Xing in distributed machine learning systems. Through this extended academic family, Russell's framing of intelligence as decision-making under uncertainty, and his emphasis on the responsible design of AI systems, has propagated through much of the contemporary research community.
The modern AI safety research community owes a substantial debt to Russell's willingness to bring concerns about advanced AI into mainstream computer science. In the 2000s and early 2010s, public discussion of long-term AI risk was dominated by figures outside the academic mainstream, and many AI researchers regarded such concerns as speculative or distracting. Russell's reputation as a co-author of the standard textbook gave him standing to argue that the field's conceptual foundations themselves needed revision and that such revision was a respectable technical research program rather than a distraction.
The technical research programs that emerged at CHAI, OpenAI, Anthropic, DeepMind, and other organisations in the late 2010s and 2020s, on inverse reinforcement learning, assistance games, reward modelling, scalable oversight, interpretability, and corrigibility, draw directly on the framing he set out in Human Compatible and in earlier work. His three principles for provably beneficial AI are routinely cited in the literature on alignment, and the King Midas example has become one of the standard ways of explaining the alignment problem to non-specialist audiences.
Russell has also been one of the more visible critics of the current race to deploy increasingly capable foundation models without adequate evaluation. He has argued that the current generation of large language models should be treated as scientific experiments rather than mature products, and that releasing them as commercial systems used by hundreds of millions of people without serious safety guarantees is irresponsible. These views have made him a regular interlocutor for governments and standard-setting bodies as the field moves toward more general capabilities and the prospect of artificial general intelligence.
Russell holds dual British and American nationality and lives in Berkeley, California. He is a chess player and has spoken about chess as one of the original test-beds for the rational-agent view of intelligence that AIMA helped establish. He is a fellow of Wadham College, Oxford, and maintains close ties to the United Kingdom AI research community.