Shane Legg
Last reviewed
May 20, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,353 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 20, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,353 words
Add missing citations, update stale details, or suggest a clearer explanation.
Shane Legg (born 1973) is a New Zealand-born machine learning researcher and entrepreneur who co-founded the artificial intelligence laboratory DeepMind in 2010 with Demis Hassabis and Mustafa Suleyman, and serves as Chief AGI Scientist at Google DeepMind.[1][2] His doctoral work at the Dalle Molle Institute for Artificial Intelligence Research (IDSIA) in Lugano, supervised by Marcus Hutter, produced the 2008 thesis "Machine Super Intelligence" and the widely cited Legg-Hutter universal intelligence measure, a mathematical formalism that defines intelligence as an agent's ability to achieve goals across a wide range of environments.[3][4] Legg is also credited, alongside Ben Goertzel and Peter Voss, with reintroducing and popularizing the term "artificial general intelligence" in the early 2000s, the name that came to label an entire subfield once Goertzel's edited volume Artificial General Intelligence appeared in 2007.[1][5] He is known for a long-running public prediction, first put forward in 2009 and made quantitative on his vetta.org blog at the end of 2011, that human-level artificial general intelligence follows roughly a log-normal distribution with a mode around 2025 and a mean around 2028, a forecast he has reaffirmed in interviews through 2023 and 2025.[6][7][8] At Google DeepMind, Legg leads the AGI Safety Council, supervises the Technical AGI Safety and Alignment team, and was acknowledged as a contributor to successive versions of the Frontier Safety Framework published in 2024 and 2025.[9][10][11]
Legg grew up in Rotorua on New Zealand's North Island and attended Rotorua Lakes High School.[1] He completed an undergraduate degree in computing and mathematical sciences at the University of Waikato in 1996, where he contributed to the WEKA machine learning toolkit project.[1] In the same year he completed a Master of Science at the University of Auckland with a thesis on Solomonoff induction, supervised by Cristian S. Calude.[1] Solomonoff induction, the algorithmic-probability-based theory of optimal sequence prediction developed by Ray Solomonoff, became the conceptual backbone of all of Legg's later theoretical work; the choice of Calude, a researcher in algorithmic information theory, as an advisor anticipated the trajectory that took him to IDSIA roughly a decade later.[1][4]
Between 1996 and the early 2000s Legg worked in a series of software development jobs at private companies, including the firm Adaptive Intelligence (later Adaptive A.I. Inc.) and the New York startup Webmind, which had been founded by Ben Goertzel to pursue commercial applications of an early AGI research programme.[1] It was during this period, around 2001 and 2002, that Legg, Goertzel, and Peter Voss settled on the phrase "artificial general intelligence" to describe systems aimed at the broad cognitive flexibility associated with human intelligence rather than narrowly trained domain solvers.[1][5] The term had been used in isolated earlier writings, but the trio's adoption of it, together with the publication in 2007 of Goertzel and Cassio Pennachin's edited collection Artificial General Intelligence, established it in common usage. In that book Legg and Hutter contributed a chapter on a formal definition of intelligence for artificial systems, an early statement of the work that would shortly appear in journal form.[5]
In 2003 Legg moved to Switzerland to pursue a doctorate under Marcus Hutter at the Dalle Molle Institute for Artificial Intelligence Research (IDSIA), an institute jointly run by the Università della Svizzera italiana (USI) and the Scuola universitaria professionale della Svizzera italiana (SUPSI) and chaired by Jürgen Schmidhuber.[3][4] At IDSIA, Legg and Hutter began the long collaboration that produced the universal intelligence formalism and an extended analysis of theoretical superintelligent agents. Legg submitted his thesis, "Machine Super Intelligence", in 2008 to the Faculty of Informatics at USI in Lugano; Hutter (by then at Australian National University) acted as supervisor, with Schmidhuber and other faculty serving on the committee.[3][4] In 2008 the thesis won the Canadian Singularity Institute for Artificial Intelligence research prize of US$10,000, awarded for outstanding contributions to the formal theory of machine intelligence.[1][4]
The central technical contribution that Legg is associated with is the universal intelligence measure that he and Hutter published over a series of papers between 2005 and 2008. The decisive presentation appeared in the article "Universal Intelligence: A Definition of Machine Intelligence", first released as a preprint on arXiv on 20 December 2007 and published in the journal Minds and Machines (volume 17, issue 4, pages 391 to 444) in 2007.[4][12] The paper carries IDSIA report number 10-07.[4]
The argument of the paper starts from a survey of informal definitions of intelligence collected from psychologists, AI researchers, and dictionaries. Legg and Hutter extract a common core: intelligence is, in their phrasing, an agent's general ability to achieve goals across a wide range of environments. They then formalise this intuition inside the framework of reinforcement learning. An agent interacts with an environment in discrete cycles: at each step the agent emits an action, and the environment returns an observation together with a scalar reward. An "intelligent" agent is one that, over many such cycles, accumulates high expected reward.[4][12]
To make the definition "universal", the authors weight performance across all computable environments by the Solomonoff prior, which assigns higher prior probability to environments with shorter algorithmic descriptions. The resulting quantity, the Legg-Hutter universal intelligence measure, is the expected reward an agent accumulates in a computable environment drawn from this prior, summed over a chosen reward range.[4][12] Two consequences follow from the construction. First, the measure is connected to Hutter's earlier AIXI framework, a theoretical agent that maximises expected reward against a Solomonoff-distributed environment and that, in the limit of infinite computational resources, attains the maximum possible universal intelligence score.[4][13] Second, the measure is only asymptotically computable: because the Solomonoff prior is uncomputable, no practical test can directly evaluate it, although a number of finite approximations have since been proposed by other researchers.[4]
Legg's PhD thesis, "Machine Super Intelligence", extends and unifies this work. The first half of the dissertation surveys philosophical, psychological, and computer-science definitions of intelligence, motivates the universal measure, and lays out its formal properties. The second half analyses AIXI as a model of optimal universal intelligence, examines the convergence of approximations to AIXI under various computational restrictions, and reflects on the implications of these theoretical results for the long-run prospects of machine superintelligence.[3] The thesis is licensed under a Creative Commons Attribution-ShareAlike licence and is hosted on Legg's vetta.org site, which functioned as his personal research blog and document repository throughout the late 2000s and 2010s.[3]
After completing his PhD, Legg spent roughly a year as a postdoctoral researcher at the Swiss Finance Institute, building probabilistic models of human decision making, before relocating to London to take a research-associate position at the Gatsby Computational Neuroscience Unit at University College London.[1] The Gatsby Unit, founded in 1998 with funding from the Gatsby Charitable Foundation, was at the time a leading hub for theoretical and computational approaches to brain function. It is where Legg first met Demis Hassabis, a fellow postdoc whose work spanned hippocampal memory and machine learning.[14]
Legg, Hassabis, and Hassabis's school friend and entrepreneur Mustafa Suleyman founded DeepMind Technologies in London in 2010 with the stated mission of "solving intelligence" through the development of artificial general intelligence in a way that benefits humanity.[2][15] The combination of academic credentials was unusual: Hassabis brought a neuroscience PhD and a track record as a video game designer, Suleyman brought experience in policy and social entrepreneurship, and Legg brought theoretical depth in universal intelligence and a long-standing engagement with the AGI research community.[2][15]
Early funding came from a mix of well-known Silicon Valley sources, including Founders Fund, Horizon Ventures, and Peter Thiel, with Thiel and his colleagues investing roughly US$2.25 million in the first major round.[15] The company recruited heavily from machine learning, neuroscience, and theoretical computer science, and built early demonstrations centred on training reinforcement learning agents to play classic Atari 2600 video games from raw pixels using the Deep Q-Network (DQN) algorithm.[16] The Atari work, published in 2013 and in a Nature paper in 2015, is widely cited as one of the trigger events that drew major industry attention to deep reinforcement learning.[16]
Google acquired DeepMind in January 2014 for an amount widely reported in the press as in the range of US$400 to US$650 million; subsequent disclosures placed the figure at roughly US$500 million.[15][17] As part of the deal Google reportedly agreed to establish an internal AI ethics board, an early example of the kind of governance structure that would later become standard at frontier labs.[17] After the acquisition, DeepMind continued to operate as a relatively autonomous subsidiary headquartered in London, producing the AlphaGo, AlphaZero, and AlphaFold systems among other research outputs, while Legg remained on the technical side of the organisation rather than its public-facing leadership.[2][16]
In April 2023, Alphabet announced that DeepMind would merge with the Brain team from Google Brain to form a unified Google DeepMind division responsible for the company's most advanced AI research, including the Gemini family of foundation models.[18] Through that reorganisation, Legg's title was clarified as Chief AGI Scientist, with explicit responsibility for the long-horizon AGI research and safety agenda of the merged unit.[9]
By his own description in interviews and in DeepMind's published materials, a large fraction of Legg's job has long been recruiting, strategy, and AGI safety leadership rather than first-author publication.[2][9] As of the mid-2020s his most visible internal responsibilities are leading the AGI Safety Council, an internal body at DeepMind that analyses AGI risks and recommends safety measures, and acting as executive sponsor for the AGI Safety and Alignment team.[9][10]
A 2017 profile noted that Legg's role then included supervising recruitment, deciding where DeepMind should focus its research efforts, and leading the company's AI safety work.[2] In a Google DeepMind blog post published on 2 April 2025 titled "Taking a responsible path to AGI", the organisation set out a framework that grouped risks from advanced AI under four headings (misuse, misalignment, mistakes, and structural risks) and described the AGI Safety Council, which Legg leads, as the body that "analyses AGI risk and best practices, making recommendations on safety measures".[9] The same document highlights specific research programmes in amplified oversight, AI Alignment via debate-style techniques, interpretability, and collaborations with external safety organisations including Apollo Research and Redwood Research.[9]
Legg has co-authored or sponsored a number of position papers from DeepMind that aim to set common terminology for the field. The most cited of these is "Levels of AGI for Operationalizing Progress on the Path to AGI", published on arXiv on 4 November 2023 by Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Legg, and presented at the International Conference on Machine Learning in 2024.[19] The paper proposes a matrixed taxonomy of AGI performance levels (Emerging, Competent, Expert, Virtuoso, and Superhuman) crossed against generality and autonomy axes, and articulates six principles for what an AGI ontology should do, including focusing on capabilities rather than processes and acknowledging that the path to AGI is itself a sequence of intermediate systems rather than a single threshold event.[19]
Legg is also one of the named contributors to the Frontier Safety Framework (FSF), DeepMind's structured approach to identifying and mitigating severe risks from frontier AI capabilities. The framework was first published in 2024 and updated to version 2.0 in a blog post dated 4 February 2025; the update introduced security-level recommendations mapped to Critical Capability Levels, a more rigorous deployment-mitigations process, and an explicit programme of work on deceptive alignment.[11] A third iteration of the FSF was subsequently described as DeepMind's most comprehensive treatment to date of risks from advanced models approaching AGI.[20] Legg is acknowledged in the author credits of these updates as having provided substantive contributions.[11]
Legg is one of the few senior researchers at a major AI lab who has consistently committed to a public, quantitative AGI timeline, and the trajectory of that timeline is by now itself a small piece of the AI-history record.
He has said in interviews that his interest in machine intelligence timelines dates back to around 1999, around the time he read Ray Kurzweil's The Age of Spiritual Machines.[7] By 2009 he was publicly stating, on his vetta.org blog and in talks, that his prediction "for the last 10 years has been for roughly human-level AGI in the year 2025".[21] The most cited statement of his view is the year-end post on his vetta.org blog titled "Goodbye 2011, hello 2012", published in late December 2011, in which Legg formalised the prediction by giving it a probability distribution: "I give it a log-normal distribution with a mean of 2028 and a mode of 2025, under the assumption that nothing crazy happens like a nuclear war."[6][21] In the same post he predicted that he would expect to see an impressive proto-AGI within the next eight years, defining the milestone informally as a system with basic vision, sound processing, motor control, and language abilities, all "essentially learnt rather than preprogrammed", and able to solve a range of simple novel problems.[6]
In a June 2011 LessWrong Q&A, Legg gave a coarser set of cumulative probabilities for human-level AGI: roughly a 10 percent chance by 2018, a 50 percent chance by 2028, and a 90 percent chance by 2050, conditional on no major civilisational disruption.[22] In the same interview he placed AI risk at the top of his list of existential threats for the twenty-first century, with engineered biological pathogens a "close second", and gave a wide range of 5 to 50 percent when asked about the probability that human extinction would follow shortly after the arrival of human-level AI.[22]
Twelve years later, in a long-form podcast interview with Dwarkesh Patel published on 26 October 2023, Legg explicitly reaffirmed the 2011 distribution. He restated the 50 percent by 2028 figure, while pointing out that what counted as AGI was itself contested and that he expected unexpected research bottlenecks along the way.[7] He proposed in the same conversation that an honest practical test of AGI would be to give a team of experts one or two months of access to a candidate system and ask them to find a single cognitive task that typical humans can do that the system fails at; if no such failure mode could be exhibited, then by his criterion the system would qualify as a minimal AGI.[23] In subsequent commentary in 2024 and 2025, Legg has emphasised that "AGI" is best treated as a field of study or a shorthand for a class of systems rather than as a sharp line, and has suggested that a full AGI, in the sense of the original DeepMind mission statement, would arrive a few years after a "minimal AGI" of the kind he expects in the late 2020s.[23][24]
The table below collects the public statements that anchor the prediction history.
| Year | Venue | Prediction |
|---|---|---|
| 2009 | vetta.org blog | Mode of human-level AGI around 2025, with sceptics likely to deny the milestone when it arrives.[21] |
| 2011 | LessWrong Q&A (June) | Cumulative probability of human-level AGI: roughly 10% by 2018, 50% by 2028, 90% by 2050.[22] |
| 2011 | vetta.org "Goodbye 2011, hello 2012" | Log-normal distribution with mean 2028, mode 2025, conditional on no civilisational disruption; expects impressive proto-AGI within 8 years.[6] |
| 2023 | Dwarkesh Podcast (October) | Reaffirms 50% by 2028; proposes one-to-two-month expert-team test for minimal AGI.[7][23] |
| 2025 | Public talks and writeups | Treats "AGI" as a field-of-study term; expects full AGI within a few years of minimal AGI.[24] |
Legg's interest in AI safety predates the contemporary alignment field and is woven through his theoretical work. In the 2011 LessWrong Q&A he framed his concern in unusually direct terms: "It's my number 1 risk for this century, with an engineered biological pathogen coming a close second."[22] He has cited the difficulty of safety research on systems that have not yet been built ("How do we make something safe when we don't properly understand what that something is or how it will work?") as a reason that conceptual and theoretical work on alignment matters even before frontier capabilities arrive.[22]
In May 2023 Legg, together with Demis Hassabis, Geoffrey Hinton, Yoshua Bengio, Stuart Russell, and several hundred other researchers and executives, signed the statement on AI risk issued by the Center for AI Safety, which read in full: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."[25] Legg's signature is consistent with the position he had taken in the 2011 Q&A and in his thesis, both of which treated the long-run safety of superintelligent systems as a first-order problem.
His preferred safety strategy, as articulated in the Dwarkesh Patel interview and in a 2025 set of talks, emphasises producing aligned behaviour rather than containment. He argues that highly capable systems cannot reliably be boxed or restricted by external mechanisms alone; instead, alignment must be built into the training process and reinforced by deliberative, multi-step "System 2" reasoning that allows a model to reflect on the ethical implications of its actions before acting.[7][24] He has supported the development of DeepMind's amplified oversight and AI-debate research programmes, scalable scalable oversight techniques, interpretability, and collaborations with external safety labs as practical instantiations of this approach.[9]
Beyond research, Legg's safety advocacy includes participation in policy-facing initiatives. He has spoken on AGI risks at the Oxford Martin School and at TED conferences, served on the nominating committee of the AI Safety Foundation, and contributed to discussions surrounding the UK Bletchley Declaration of November 2023 in his capacity as a Google DeepMind co-founder, although he has generally kept a relatively low public profile compared with Hassabis and Suleyman.[2][26]
Legg was appointed Commander of the Order of the British Empire (CBE) in the 2019 Birthday Honours "for services to the science and technology sector and to investment".[27] In September 2023 he was named to Time magazine's inaugural Time100 AI list of the most influential people in artificial intelligence.[28] He is a frequent invited speaker at academic and industry venues, including a TEDAI San Francisco talk and lectures hosted by USI Lugano on the trajectory from theoretical AGI research to industrial deployment.[29]
Legg occupies an unusual position in the modern AI landscape. He is one of a small number of researchers whose work straddles the foundational theory of intelligence (the Legg-Hutter formalism), the founding of a major industrial AI laboratory (DeepMind), and the practice of long-term AI safety at the scale of a frontier model developer (the AGI Safety Council and the Frontier Safety Framework).[4][9][11] The phrase "artificial general intelligence" itself entered mainstream use largely because of his early advocacy alongside Ben Goertzel and Peter Voss, and the Goertzel-Pennachin volume to which he contributed gave the field a name and a research identity well before deep learning had matured into a credible technical path.[1][5]
His public predictions are equally significant. The 2011 vetta.org post is one of the earliest quantitative AGI forecasts from a serious researcher that gave a specific probability distribution rather than a hand-waved date, and the persistence of the same headline figure across more than a decade and a half has made it a reference point for subsequent forecasting exercises.[6][7] Whether or not the 2028 median proves accurate, the post has shaped the way the field talks about AGI timelines, including the now-common practice of stating a distribution and a conditional assumption rather than a single year.[21]
Aspects of Legg's research programme have attracted reasoned criticism. The Legg-Hutter universal intelligence measure, while widely cited and influential, has been criticised on the grounds that its reliance on the Solomonoff prior makes it uncomputable in practice, that it inherits the reward-maximisation framing of reinforcement learning (which not all researchers regard as a sufficient model of intelligence), and that any practical approximation of the measure depends sensitively on the choice of reference universal Turing machine.[4][30] Researchers including Hernandez-Orallo have proposed alternative or modified measures intended to address some of these issues.[30]
His AGI predictions have also drawn pushback from both directions: from sceptics who view a 50 percent median by 2028 as wildly optimistic given the current state of multimodal reasoning and embodied capabilities, and from short-timeline advocates such as parts of the LessWrong community who regard his probability mass on the late 2020s and early 2030s as too conservative given the post-2022 capability gains in large language models.[22][24] Legg himself has been explicit that he expects unforeseen research obstacles and that AGI is most useful as a class of systems rather than a single sharp threshold, a stance that some critics regard as a sign that the term may be losing precision precisely as it becomes more politically charged.[7][23]
A separate critique, levelled by writers in the broader AI policy debate, is that holding a senior AGI safety role inside a frontier developer creates structural tensions with a stated long-term concern about existential risk from AI. Legg has defended his position by arguing that the most rigorous safety work has to happen close to the systems being built, and that his role inside Google DeepMind allows him to translate theoretical concerns about misalignment and deceptive alignment into specific evaluations and deployment policies through bodies such as the AGI Safety Council and the Frontier Safety Framework.[9][11]
Legg's intellectual influences and collaborators include Marcus Hutter, the originator of AIXI and his PhD supervisor, and Jürgen Schmidhuber, the IDSIA director on whose lab the universal intelligence programme was based.[3][4] His DeepMind co-founders Demis Hassabis and Mustafa Suleyman took complementary roles: Hassabis as scientific and public-facing lead and CEO of Google DeepMind, Suleyman as policy and applied-AI lead until his departure to co-found Inflection AI and later to lead Microsoft's consumer AI organisation.[15][18]
Within the contemporary AI safety field, Legg's combination of high probability mass on near-term AGI and senior insider status places him in close conceptual proximity to figures such as Geoffrey Hinton, Yoshua Bengio, and Stuart Russell, all of whom co-signed the May 2023 Center for AI Safety extinction-risk statement.[25] His preferred research strategy (deliberative System-2 reasoning, scalable oversight, interpretability, and outside-lab collaborations with Apollo Research and Redwood Research) overlaps substantially with the wider alignment research agenda set out under headings such as Superalignment and scalable oversight.[9]