- Co-Founder and Chief AGI Scientist, Google DeepMind23
Shane Legg is a co-founder of DeepMind, which Demis Hassabis wrote was founded in 2010 along with Legg and Mustafa Suleyman, and is Co-Founder and Chief AGI Scientist at Google DeepMind. Google DeepMind's responsibility page says Legg leads the AGI Safety Council, which works with the Responsibility and Safety Council to safeguard the lab's processes, systems and research against extreme risks from future AGI systems. Legg is a co-author of Universal Intelligence: A Definition of Machine Intelligence (2007), of Levels of AGI for Operationalizing Progress on the Path to AGI (ICML 2024), of the April 2025 post Taking a responsible path to AGI, and of the 2026 report From AGI to ASI.123456
At a glance
- Current role
- Co-Founder and Chief AGI Scientist
- Company
- Google DeepMind
- Sector
- Company
- Work listed
- 6
- Sources cited
- 11
- Profile reviewed
Career
Shane Legg co-founded DeepMind in 2010 with Demis Hassabis and Mustafa Suleyman, and Hassabis's 2019 post describes Legg as the company's Chief Scientist. Google DeepMind now lists Legg as Co-Founder and Chief AGI Scientist.12
At Google DeepMind, Legg leads the AGI Safety Council, which analyzes AGI risk and best practices and recommends safety measures. The council works closely with the Responsibility and Safety Council to guard the lab's processes, systems and research against extreme risks from future AGI systems.23
Research
With Marcus Hutter, Legg published two 2007 papers on what intelligence means: a survey of about seventy informal definitions, and Universal Intelligence, which starts from well-known informal definitions and works toward a formal measure for machines. Legg is also among the authors of DeepMind's 2015 Nature paper on human-level control through deep reinforcement learning.748
Legg co-authored several safety papers from DeepMind, including Deep reinforcement learning from human preferences (2017), AI Safety Gridworlds (2017), a suite of environments for testing safety properties of agents, and Scalable agent alignment via reward modeling (2018), which outlines a research direction for the agent alignment problem.91011
AGI frameworks
Legg is a co-author of Levels of AGI (ICML 2024), which proposes levels of performance, generality and autonomy for comparing systems and measuring progress. Legg also co-wrote Google DeepMind's April 2025 post Taking a responsible path to AGI and the June 2026 report From AGI to ASI, which looks at how AI might keep developing after human-level AGI.536
Work and contributions
Levels of AGI for Operationalizing Progress on the Path to AGI Published research
Co-author of the ICML 2024 paper, which proposes a framework of levels of AGI performance, generality and autonomy for comparing models, assessing risks and measuring progress.5
From AGI to ASI Published research
Co-author of the June 2026 Google DeepMind report on routes from human-level AGI to superintelligence.6
Universal Intelligence: A Definition of Machine Intelligence Published research
Co-author, with Marcus Hutter, of the 2007 paper that proposes a formal measure of machine intelligence.4
AGI Safety Council (Google DeepMind) Proprietary product
Leads the AGI Safety Council, which analyzes AGI risk and best practices and makes recommendations on safety measures, according to Google DeepMind.23
Human-level control through deep reinforcement learning Published research
Co-author of DeepMind's 2015 Nature paper on an agent that learns control policies directly from high-dimensional inputs through deep reinforcement learning.8
Deep reinforcement learning from human preferences Published research
Co-author of the 2017 paper that trains reinforcement learning agents from non-expert human preferences between pairs of trajectory segments instead of access to the reward function.9
Timeline
- Released A Collection of Definitions of Intelligence with Marcus Hutter.7
- Released Universal Intelligence, a formal definition of machine intelligence, with Marcus Hutter.4
- Co-founded DeepMind with Demis Hassabis and Mustafa Suleyman.1
- Co-authored DeepMind's Nature paper on human-level control through deep reinforcement learning.8
- Co-authored Deep reinforcement learning from human preferences.9
- Co-authored AI Safety Gridworlds, a suite of environments for testing safety properties of agents.10
- Co-authored Scalable agent alignment via reward modeling, a proposed research direction.11
- Co-authored Levels of AGI for Operationalizing Progress on the Path to AGI, which appeared at ICML 2024.5
- Co-wrote Google DeepMind's post Taking a responsible path to AGI.3
- Co-authored the Google DeepMind report From AGI to ASI.6
Other documented roles
Roles the sources document that are not confirmed as current.
Where they work
- Google DeepMind 64 people in the wiki
Related people
- Sven GowalAlso at Google DeepMind
- David StutzAlso at Google DeepMind
- Roy FrostigAlso at Google DeepMind
- Lila IbrahimAlso at Google DeepMind
- Anca DraganAlso at Google DeepMind
- John LangfordMachine-learning theory
- Mido AssranMachine-learning theory
- James GlasbrennerSafety and alignment
- Geoffrey HintonMachine-learning theory
- Jack ClarkSafety and alignment
Suggest a correction or removal
Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.
References
11 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026
- From unlikely start-up to major scientific organisation: Entering our tenth year at DeepMind
- Responsibility & Safety — Google DeepMind
- Taking a responsible path to AGI
- Universal Intelligence: A Definition of Machine Intelligence (arXiv 0712.3329)
- Levels of AGI for Operationalizing Progress on the Path to AGI — Google DeepMind
- From AGI to ASI — Google DeepMind
- A Collection of Definitions of Intelligence (arXiv 0706.3639)
- Human-level control through deep reinforcement learning (Nature 518, 529–533)
- Deep reinforcement learning from human preferences (arXiv 1706.03741)
- AI Safety Gridworlds (arXiv 1711.09883)
- Scalable agent alignment via reward modeling: a research direction (arXiv 1811.07871)