Google has spent the better part of a decade insisting it takes AGI seriously. On Wednesday it built a place to argue about it, and then took care to make sure nobody could hold the company to anything said there.
The DeepMind Institute, launched September 16, is a publishing platform for essays on artificial general intelligence, directed by DeepMind co-founder Shane Legg, DeepMind chair and Alphabet chief scientist Demis Hassabis, and James Manyika, Google's president of research, labs, technology and society. Legg, whose title at DeepMind is Chief AGI Scientist, doubles as managing editor. The site's charter promises contributions from Google DeepMind, Google and the wider research community, and every piece carries the same caveat: the ideas belong to the authors and, in the institute's words, "should not be read as Google's official view."
That disclaimer is the most interesting design choice on the site. It lets DeepMind's most senior people float proposals with real policy consequences, including a coordinated industry slowdown, without Alphabet's lawyers or lobbyists having to defend any of them.
The inaugural collection runs to five pieces: an introduction by the three directors, plus four essays. Rohin Shah and Anca Dragan, both DeepMind safety researchers, argue that the shrinking ability to read a model's step-by-step reasoning is a choice rather than a law of nature, and suggest regulators could cap what they call "opaque serial depth," the amount of sequential computation a model does without emitting a legible trace. Stephen Cave contributes principles for what he terms a new utopianism. Hassabis republishes his July proposal for a U.S.-led frontier AI standards body, under which developers would initially submit models voluntarily up to 30 days before release, with the evaluations eventually becoming mandatory and built on undisclosed "held-out" tests so labs cannot train to the exam. Hassabis writes the framework could be "ratcheted up if the seriousness of the situation demands," a phrase that does a lot of work given the antitrust suit currently alleging frontier labs coordinated to slow releases.
The fourth essay is the one with the most numbers. Julian Jacobs, a DeepMind research scientist, and Alex Imas, its Director of AGI Economics, scored 11 economic policies for an AGI-disrupted labor market using literature reviews, public surveys and 51 AI agent raters calibrated to the views of 51 real economists. Universal basic capital, a scheme that gives citizens an ownership stake in AI-driven growth, scored highest on agency at 76.3 out of 100 but second lowest on feasibility at 33.1, ahead only of a federal jobs guarantee. Universal basic income fares worse in their telling: "an expensive and blunt instrument that may fail to concentrate sufficient relief where it is needed most." Their survey found 85% of Americans back publicly funded retraining, 72% back expanded unemployment insurance, and 54% back universal basic capital. The authors sort their recommendations into three scenarios, from mild disruption (expand the Earned Income Tax Credit, employer-led retraining) to a full decoupling of labor and capital, at which point the capital backstop kicks in if labor's share of GDP falls for a sustained period.
The introductory essay sets the tone the directors want. Google, DeepMind and outside researchers "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, Legg and Manyika write. The institute exists, they add, "because broad-based intellectual discussion and debate are required to arrive at a consensus about how to address the challenges and opportunities we face as a society."
Timing is not incidental. The launch came four days after Anthropic CEO Dario Amodei's September 12 essay calling on the industry to "pace" frontier development, slowing but not pausing releases. Legg told the Financial Times that Amodei's proposal was "interesting directionally" and "worth considering," and separately said it was premature to declare AGI achieved, a pointed response to OpenAI's September 4 AGI claim and Jensen Huang's September 7 assertion that it had already arrived. Legg said he remains comfortable with his long-standing forecast of a 50% chance of "minimal" AGI by 2028. The institute's own launch essay hedges the other direction, conceding today's systems "still fail at some basic tasks" while adding "we expect those gaps to be closed soon."
Manyika, in an interview with Axios, framed the exercise as a global one. "Humanity has to be a part of this. I don't know how we think about safety and responsibility on a country-by-country basis. This is a global technology that will be used by everybody," he said.
Why It Matters
The obvious comparison is to what Google's rivals already run. OpenAI has a policy shop that publishes economic blueprints and an "Economic Index" tracking how people use ChatGPT for work; Anthropic publishes its own Economic Index and used Amodei's personal essay, not a corporate white paper, to propose slowing the frontier. DeepMind's version is structurally different: it is an editorial vehicle rather than a lobbying arm, and the disclaimer is the feature. When Hassabis proposes mandatory pre-release evaluation and the possibility of a coordinated slowdown, that is a proposal from an essayist, not a commitment from Alphabet, which sells Gemini and is a defendant in the same regulatory environment it is describing.
There is a less charitable read. DeepMind lost Bilal Chughtai from its AGI safety team in July; he posted a public warning this week, and the lab's own 145-page AGI safety paper from April 2025 has spent 17 months as a document without an institutional home. A think tank with a Chief AGI Scientist as editor is a credible recruiting pitch to safety researchers deciding between labs, and a credible venue for the kind of policy proposals Google has historically declined to attach its name to. It is also, frankly, cheap: five essays and a website, launched the week the safety conversation shifted from statements of concern to concrete mechanisms.
What the institute has not done is commit to anything measurable. There is no stated cadence, no editorial board beyond the three directors, no disclosure of how outside authors are selected or paid, and no mechanism for the debate it promises to actually change a Google product decision. The economic essay's most striking finding, that the policy with the highest agency score is nearly the least feasible, is presented as an argument for more discussion. That is what think tanks do. It is not what regulators, or the lawmakers in the UK who this week asked for a superintelligence ban, are asking for.
What to Watch
The test is whether the second batch of essays includes anyone who disagrees with the first. The directors promised contributors who "will not always agree"; the launch lineup is three DeepMind staff essays, one from an outside academic, and a republished op-ed by the chair. Watch for whether Amodei, OpenAI's policy team or independent critics like Chughtai are invited in, whether Hassabis's 30-day voluntary evaluation window shows up in any actual U.S. legislative text this fall, and whether Legg's "worth considering" on pacing hardens into a DeepMind release policy or stays, as everything on the site does, just one author's view.
"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."— Demis Hassabis, Shane Legg and James Manyika, Directors, DeepMind Institute