Google DeepMind co-founder warns AI progress must not outrun safety

Google DeepMind co-founder Shane Legg has launched a new institute focused on the implications of artificial general intelligence (AGI), while warning that technical progress should be matched by stronger safeguards. Legg says AGI remains unproven, but continues to stand by his forecast of a 50 per cent chance of minimal AGI by 2028.

The race to develop increasingly capable AI needs to be matched by equally serious work on safety, according to Google DeepMind co-founder Shane Legg, who has established a new institute to examine what the arrival of artificial general intelligence could mean for society.

Legg's intervention comes as AI researchers and executives increasingly debate whether systems capable of improving their own capabilities could accelerate technological development beyond the ability of existing safeguards to keep pace. Today's advanced agents can already write software, carry out tasks with limited supervision and contribute to AI research, although the extent to which these capabilities could lead to recursive self-improvement remains uncertain.

The new DeepMind Institute is intended to widen discussion around AGI beyond the technical development of AI systems. According to the Financial Times, its work will examine areas including science, education, economics, public policy, philosophy and human flourishing.

Google DeepMind uses AGI to describe an AI system possessing cognitive abilities comparable with those of the human brain. Legg, however, cautioned against assuming that such a milestone has already been reached.

Recent claims by senior figures at companies including Nvidia and OpenAI that AI may be approaching or achieving forms of AGI have not convinced Legg that the threshold has been crossed.

He considers it too early to declare AGI a reality, according to the Financial Times. The distinction matters because there is no universally accepted test for AGI, leaving disagreement over what capabilities would qualify a system as genuinely general rather than highly advanced within particular areas.

Legg nevertheless continues to expect substantial progress in the near term. He remains comfortable with his earlier prediction that there is a 50 per cent chance of achieving what he calls “minimal” AGI by 2028.

That forecast does not amount to a claim that AGI will arrive by that date. Instead, it reflects Legg's assessment of the probability of reaching a relatively basic form of general intelligence within that timeframe.

The institute's opening publications are designed to broaden the debate. Its first three essays examine AI safety, economic policy and the principles behind what it describes as a “new utopianism”.

Debate grows over pace of AI development

Legg's comments come as companies developing frontier AI systems confront growing questions about how quickly increasingly autonomous models should be released.

Anthropic chief executive Dario Amodei has argued that the development of frontier systems should slow rather than stop. Legg described that approach as “interesting directionally” and said it was “worth considering”, according to the Financial Times.

The discussion reflects a wider concern that technical capability could advance faster than institutions can develop effective methods for monitoring and controlling powerful systems. AI agents are already being used for autonomous software development and other multi-step tasks, providing researchers with an indication of how much more independent future systems might become.

Legg's new institute is therefore positioned less as a prediction about when AGI will arrive than as an effort to examine the consequences if it does. Its remit spans questions about how such technology could affect economies, public institutions and human wellbeing.