Shane Legg on Defining AGI, Safety and the Future of Work

In a December 2025 Google DeepMind podcast, Hannah Fry asks co-founder and chief AGI scientist Shane Legg how he defines artificial general intelligence and what its development might mean. His timelines and social predictions are opinions, not established outcomes.

Legg calls minimal AGI an artificial agent able to perform the kinds of cognitive tasks people can typically perform. He says systems at the time show impressive capabilities but also important gaps, so he does not consider the threshold reached.

He points to strong language and broad knowledge alongside weaknesses in reliability and visual reasoning. A model may miss an implication of perspective in an image even while answering other difficult questions well.

Legg expects closing the long tail of weaknesses to take work across multiple areas. He discusses better data, learning methods and memory mechanisms as possible contributors rather than naming one guaranteed solution.

He distinguishes minimal AGI from a stronger “full AGI” able to match a wider range of exceptional human cognitive achievements. He uses artificial superintelligence for capabilities that go beyond human cognition, while acknowledging contested definitions.

Legg recalls helping popularize AGI as a term for the older research ambition of broadly capable thinking machines, rather than systems specialized in one task. He notes that an earlier use of the term also existed.

He says different communities use AGI to mean different thresholds, economic effects or types of systems. Without a shared definition, apparently conflicting timeline predictions can refer to different things.

Fry brings up alternative tests, such as performing tasks in an unfamiliar kitchen or making money. Legg prefers a broad range of cognitive evaluations over a single dramatic challenge.

For his proposed threshold, failures on ordinary cognitive tasks matter even if the system excels elsewhere. He wants many tests and adversarial attempts to uncover gaps, although any finite test set has limits.

Legg thinks everyday conversation may eventually focus less on the AGI label and more on what a particular system can do. Before then, he considers multidimensional capability assessments important for understanding both strengths and weaknesses.

Fry asks whether an AI could reason ethically about its own actions. Legg says this is not solved and outlines a research direction in which a system reasons through a situation, possible actions and their consequences.

He contrasts rigid rules with context-sensitive reasoning: a rule may be useful in general but can conflict with another obligation in an unusual situation. He argues that a system must reason about such tradeoffs rather than apply slogans mechanically.

Fry asks how machines can be grounded in human ethics when values vary across people and cultures. Legg acknowledges that disagreement and discusses learning from human feedback and experience as part of the challenge.

On safety, he argues for layers: testing, monitoring, interpretability and controls on real-world interaction. He says no single method makes a complex system perfectly safe.

The conversation considers the difference between accidental harm and deliberate harmful behavior, and whether deployment should be limited until a system meets a safety threshold. Legg describes this as a difficult judgment.

He describes evaluating dangerous capabilities, including hacking, and says serious risks could lead to a decision not to release a model. The transcript is a description of his stated process, not an audit of an individual release.

Fry asks about AI consciousness. Legg says there is no agreed answer and imagines that future systems could produce disagreement about whether they have subjective experience. Talking about consciousness alone would not settle the issue.

They discuss how quickly a full AGI, if achieved, might move toward capabilities beyond humans. Legg says the speed of such a transition is a major open question rather than a known schedule.

Legg compares biological and artificial systems in terms of energy, memory and communication speed. He argues that physical differences could permit some machine capabilities beyond human limits, while the exact consequences remain uncertain.

He says the possibility of very capable systems motivates research into safety reasoning and oversight. Strong performance on a benchmark alone would not answer whether a system behaves appropriately in novel situations.

The conversation shifts to economic organization. If machines perform more cognitive labor, Legg says society may need to reconsider how work, access to resources and wealth distribution operate.

He wants researchers beyond computer science—including economists, lawyers, psychologists and other disciplines—to examine their own institutions and prepare for possible changes.

Fry asks about uneven disruption along the way. Legg offers a hypothetical reduction in staffing for some software tasks as an illustration; it is a scenario, not a measured forecast for a specific firm.

He suggests that work performed entirely remotely through a computer may be more directly exposed to cognitive automation than tasks requiring physical presence. This rule of thumb has many exceptions and does not determine individual outcomes.

Legg cautions against assuming that current limitations will remain fixed, while Fry presses on the difficulty of adjusting to rapid changes. Both the pace of progress and its labor effects are uncertain.

He revisits his long-standing timeline for minimal AGI and notes that a stronger form could take additional years. These are his personal forecasts and depend on his particular definitions.

Fry asks whether the prospect feels bleak. Legg instead emphasizes the possibility of greater abundance, but says society must consider how the benefits are shared and what allows people to flourish.

He wants the discussion to include philosophy, economics, psychology and ethics, not only technical capability. The interview closes with an invitation to consider choices society may face as systems improve.