Google and Google DeepMind launched the “DeepMind Institute” on September 17, 2026, with the aim of broadening the discussion around artificial general intelligence (AGI) and highlighting differences between the positions of Google, Google DeepMind, and researchers around the world. The institute is led by Shane Legg, DeepMind’s co-founder, alongside James Manyika, a Google executive, and Demis Hassabis, the head of Google DeepMind, while Legg also serves as managing editor.
The institute presents itself as a space for showcasing viewpoints that may not always agree and may change as new data and information emerge in a rapidly evolving field. The move comes at a time when the discussion of AI safety is shifting from general statements to more specific proposals concerning disclosure, external audits, and slowing the development of advanced models if appropriate safeguards do not keep pace.
Four Angles on the Discussion
The institute’s opening collection includes four articles addressing policies for dealing with potential economic disruptions that AGI may cause, preserving the understandability of models’ reasoning for humans, principles for supporting human flourishing, and a framework for evaluating leading artificial intelligence models.
In one article, DeepMind safety researchers Rohin Shah and Anca Dragan warn that the declining ability of researchers to see and verify a model’s reasoning steps is not inevitable. As architectures develop that make more capable models harder to monitor, the researchers call on developers and regulators to address the safety-related trade-offs directly.
Among the options they propose are limiting “opaque serial depth,” meaning the amount of sequential computation a model can perform without producing a readable reasoning trace, or requiring developers to demonstrate that less transparent systems remain monitorable to the same extent.
A U.S. Framework for Evaluating Models
In another article, Hassabis proposes establishing a U.S.-led body for advanced AI standards to evaluate the most sophisticated models. Under the proposal, developers would voluntarily submit their models for review up to 30 days before release, with passing the tests later becoming a condition for deploying leading models in the United States if the evaluation system proves effective.
The body would initially design evaluation tests in consultation with AI companies, then move to independent tests not disclosed in advance, or what the framework calls “held-out tests,” to prevent laboratories from optimizing their models for known tests. The proposal also mentions the possibility of gradually tightening measures, eventually reaching a coordinated slowdown among developers of advanced models if the seriousness of the situation requires it.
Why Does This News Matter?
The initiative’s practical significance lies not only in creating a new publishing platform, but also in bringing together Google’s and Google DeepMind’s discussions of issues that could affect how models are developed and evaluated. The articles place transparency, monitorability, and independent auditing within operational questions, rather than treating them solely as general principles.
Nevertheless, the institute is not an independent regulatory body, and the proposed evaluation framework begins voluntarily and has not become an effective policy. Questions such as the body’s independence, the nature of undisclosed tests, and the criteria for determining when development needs to be slowed remain open within the limits of what was stated in the article.