Google DeepMind Institute Proposes AGI Tests and Possible Slowdown
Google DeepMind's new institute widens the AGI debate with four essays on model transparency, frontier tests, human flourishing and AI slowdown proposals.
Summary
Google and Google DeepMind researchers launched the DeepMind Institute on September 16, 2026, to surface disagreements about artificial general intelligence across the companies and global research community. Directors are DeepMind co-founder Shane Legg, who is also managing editor, Google executive James Manyika, and Google DeepMind chair Demis Hassabis. Its inaugural four essays address economic policies for potential AGI disruption, human-readable model reasoning, principles for human flourishing, and frontier AI evaluation.
Google DeepMind safety researchers Rohin Shah and Anca Dragan argue that diminishing access to models' step-by-step reasoning is not inevitable, despite architectures becoming harder to monitor. They propose limiting opaque serial depth, the sequential computation performed without readable reasoning, or requiring proof that less transparent systems remain equally monitorable. Hassabis proposes a U.S.-led standards body offering voluntary reviews up to 30 days before release, with passing evaluations potentially becoming mandatory for U.S. deployment. Initial tests would be designed with AI companies, then replaced by independent, undisclosed held-out assessments to deter gaming. Oversight could escalate to a coordinated slowdown among frontier developers if risks intensify. The proposals follow industry leaders' endorsement this week of elements of Anthropic CEO Dario Amodei's call to pace frontier AI development.
Positives
- Four inaugural essays turn AGI safety concerns into proposals covering economic disruption, transparent reasoning, human flourishing and model evaluation.
- Reviews up to 30 days before release could expose frontier model weaknesses before public deployment.
- Independent held-out tests could prevent AI laboratories from optimizing models around known evaluation criteria.
- Rohin Shah and Anca Dragan propose preserving monitorability through limits on opaque serial depth or equivalent safety demonstrations.
Risks & concerns
- New AI architectures are making the most powerful models' step-by-step reasoning harder to inspect.
- The proposed frontier AI reviews would initially be voluntary, leaving participation to developers.
- Assessments designed with AI companies could lack independence until undisclosed held-out tests replace them.
- A coordinated slowdown may become necessary if safeguards fail to keep pace with frontier AI development.