Google DeepMind Launches Institute to Widen AGI Debate; Hassabis Proposes US-Led Standards Body

Google DeepMind has launched the DeepMind Institute to broaden the global debate on artificial general intelligence, with a proposal from Demis Hassabis for a US-led standards body that could require safety testing before frontier AI models are deployed.

Google DeepMind has launched a new institute dedicated to widening the global conversation around artificial general intelligence, bringing together researchers from Google, DeepMind and the broader scientific community to examine the technology's potential impact on society, economics and human flourishing.

The DeepMind Institute is led by Shane Legg, DeepMind's co-founder and chief AGI scientist, alongside Google senior vice president James Manyika and DeepMind chair Demis Hassabis. In its launch essay, the directors acknowledged that "they will not always agree, and they will likely change their minds, as more data and information comes to light at the fast-moving frontier."

Hassabis Proposes US-Led Standards Body

A central proposal in the institute's first collection of essays comes from Hassabis, who outlines a framework for a US-led frontier AI standards body. Under the proposal, developers would initially submit their most advanced models voluntarily for review up to 30 days before release. Once the evaluation system proved effective, passing its tests could become a mandatory requirement for deploying frontier models in the United States.

The body would initially design assessments in consultation with AI companies, but would eventually develop independent, undisclosed evaluations what Hassabis calls "held-out" tests to prevent labs from tailoring their models to known criteria. Hassabis noted that the framework could be "ratcheted up if the seriousness of the situation demands," potentially including a coordinated slowdown among frontier developers.

Reasoning Transparency in Focus

Another essay, by DeepMind safety researchers Rohin Shah and Anca Dragan, argues that AI's shrinking window of transparency is not inevitable. The authors say developers and regulators should confront safety trade-offs directly, potentially limiting "opaque serial depth" the amount of sequential computation a model can perform without producing a readable reasoning trace.

Legg on Amodei's Slowdown Call

Shane Legg told the Financial Times that AI capabilities must not outrun safety controls, calling Anthropic CEO Dario Amodei's call to slow frontier model releases "interesting directionally" and "worth considering."  Legg also said it was premature to declare AGI achieved, despite recent claims by Nvidia's Jensen Huang and OpenAI. He maintained his long-held forecast of a 50 per cent chance of "minimal" AGI by 2028.

Economic Policy for an AGI Era

The institute also published an essay on economic policy for AGI, authored by economists Julian Jacobs and Alex Imas, which rates 11 policy responses to potential labour disruption. The authors identify "least-regret" options including expanded unemployment insurance and a negative income tax, while calling universal basic income "an expensive and blunt instrument."

Here is the question this launch raises. Google DeepMind has created an institute to widen the AGI debate and proposed a US-led standards body that could mandate safety testing before deployment. When the company building the technology also proposes the rules governing it, is this genuine stewardship or a strategic move to shape regulation before it is imposed?

FAQ:

Q: What is the DeepMind Institute?
A: The DeepMind Institute is a new platform launched by Google DeepMind for research and debate on artificial general intelligence. It is led by Shane Legg, James Manyika and Demis Hassabis.

Q: What did Demis Hassabis propose regarding AI regulation?
A: Hassabis proposed a US-led frontier AI standards body that would initially review models voluntarily up to 30 days before release, with the potential to become mandatory. It would eventually develop independent, undisclosed evaluations.

Q: What did Shane Legg say about slowing AI development?
A: Legg called Dario Amodei's call to slow frontier model releases "interesting directionally" and "worth considering," while saying details need to be worked out.

Q: Did Google DeepMind declare that AGI has been achieved?
A: No. Shane Legg said it was premature to declare AGI achieved despite claims by Nvidia's Jensen Huang and OpenAI. He maintained a 50 per cent chance of "minimal" AGI by 2028.

Q: What economic policies were proposed for an AGI era?
A: An essay by Julian Jacobs and Alex Imas rated 11 policies, identifying expanded unemployment insurance, a negative income tax and a universal basic capital backstop as "least-regret" options, while calling UBI "expensive and blunt."