The DeepMind Institute is a new research foundation announced by researchers at Google and Google DeepMind, aiming to broaden and deepen the conversation around artificial general intelligence (AGI). The foundation comes at a critical time for the industry, as the conversation about AI safety shifts from general statements of concern to specific proposals for policy, transparency, and regulation. The establishment of the DeepMind Institute marks a new phase in the public engagement of major AI companies with the challenges posed by the development of advanced models.
The foundation’s board of directors includes three prominent figures in the field: Shane Legg, co-founder of DeepMind, James Manyika, a senior executive at Google, and Demis Hassabis, president of Google DeepMind. Legg will serve as editor-in-chief, suggesting that the foundation will also serve as a platform for publishing academic and policy texts. The choice of these three individuals is not accidental: all three have played a key role in shaping Google’s AI strategy in recent years.
See also: Google DeepMind: Employees against military AI contracts
According to the official announcement, the DeepMind Institute will not function as a mouthpiece for a single corporate line, but as a space for open dialogue where different views will be highlighted — even if they contradict each other. “They will not always agree, and will likely change their minds as new data and information come to light on this rapidly evolving front,” the announcement says. This approach is notable, as big tech companies rarely publicly encourage internal dissent.
DeepMind Institute: The four opening essays on AGI
The first collection of papers from the DeepMind Institute includes four essays covering a wide range of topics. The first is about economic policies to manage the potential disruption that AGI will bring to the job market and the economy in general. The second examines the preservation of human - readable model reasoning — the ability to observe and understand how an AI system reaches a decision. The third essay develops principles for human well-being in the age of artificial intelligence , while the fourth proposes an evaluation framework for the most advanced AI models worldwide.
Of particular interest is the essay by DeepMind security researchers Rohin Shah and Anca Dragan , who argue that the shrinking “window of transparency” in modern AI models is not inevitable. As new architectures make more powerful models harder to monitor, the authors argue that developers and regulators must address security compromises head on . This could mean limiting the so-called “opaque serial depth” — the amount of sequential computations a model can perform without producing a human-readable trace of its reasoning.
Shah and Dragan also propose requiring developers to demonstrate that less transparent systems remain just as observable as their predecessors. This proposal has important implications for how future AGI models are designed and developed, as it sets an explicit technical criterion for transparency as a prerequisite for development. This discussion is crucial, especially as reasoning models —such as OpenAI ’s o 3 or Google ’s Gemini 2.5 Pro —execute increasingly complex chains of thought that are difficult to externally monitor.
Hassabis and DeepMind Institute: Proposal for an international AGI
The most ambitious of the four essays is by Demis Hassabis, who proposes the creation of a U.S. standards body for frontier AI — that is, for the world’s most advanced artificial intelligence models. Under the framework he outlines, developers would initially voluntarily submit their models for review up to 30 days before release. Once the review system proves effective, passing the tests could become a mandatory requirement for frontier models to be developed in the United States.
See also: Google DeepMind opens its first AI campus in Seoul
Hassabis predicts that the agency will initially design the assessments in collaboration with AI companies, but will eventually develop independent, unpublished tests — so-called “held-out” assessments — to prevent models from being adjusted to known tests. This proposal addresses a critical problem that has emerged in the industry: so-called “benchmark gaming”where companies train their models specifically to do well on known tests, without reflecting any real improvement in their capabilities or safety.
Hassabis doesn’t rule out more drastic measures: the framework could be “tightened” if the severity of the situation warrants it, including a coordinated slowdown among developers frontier AI. This reference to a potential slowdown is particularly significant, coming from one of the industry’s most influential leaders and echoing similar calls from Dario Amodei, CEO of Anthropic, who has publicly called for “regulating” the pace of frontier AI.
The establishment of the DeepMind Institute is part of a broader climate of change in the artificial intelligence industry. In recent years, the discussion about the safety of AGI has often been abstract and theoretical. Now, however, with models becoming increasingly powerful and penetrating critical sectors such as health, education, defense and economics, the need for specific governance structures is becoming imperative. Industry leaders, regulators and researchers around the world are seeking frameworks that balance innovation with safety.
See also: Google DeepMind: How do hackers breach AI Agents?
In Europe, the European Union ’s Artificial Intelligence Act ( AI Act ) has already set a legal framework for regulating AI systems according to the risk they pose. However, Hassabis’ proposal for a US rating body could complement or even compete with the European model, creating an international regulatory landscape that will also affect Greek companies and organizations that use or develop AI systems. Greece, as a member of the EU, should closely monitor these developments, as they directly affect the country’s digital transformation strategy.
