The DeepMind Institute launched on September 16, 2026, giving Google DeepMind a new forum for research and public debate about artificial general intelligence (AGI), its risks and its economic consequences. The institute is directed by Shane Legg, Demis Hassabis and James Manyika, with Legg serving as managing editor.
The launch does not say that AGI has arrived. Instead, the institute defines AGI as a system displaying all the cognitive capabilities of the human brain while acknowledging that current systems still fail at some basic tasks. That distinction matters as the industry argues over increasingly capable models and increasingly urgent safety questions.
A research forum, not a single AGI doctrine
The DeepMind Institute brings together contributors from Google DeepMind, Google and the wider research community. Its remit spans AGI safety, governance, reasoning transparency, economic policy, global access, science, education, philosophy, society and human flourishing.
The institute’s directors say contributors may disagree and revise their positions as evidence develops. The essays represent their authors’ views rather than an official Google position. In practice, that makes DMI less like a body issuing one definition of the future and more like a platform for examining competing technical and social possibilities.
The launch materials also identify cybersecurity, biorisks and possible loss of control in self-improving systems as significant concerns. Those risks sit alongside more immediate questions about work, education, misinformation and access to advanced AI.
What the institute says AGI means
Under DMI’s definition, AGI would have the full range of cognitive capabilities associated with the human brain. That is a broader threshold than simply producing fluent text, solving selected problems or performing well on a collection of benchmarks.
Legg also said that GPT-6 Astra does not yet satisfy OpenAI’s own stated criterion of performing most economically valuable cognitive tasks. The two definitions are related but not identical, so claims about whether a system has reached AGI depend partly on which threshold is being applied.
The practical answer is therefore less dramatic than many AI headlines suggest: DMI has not announced the arrival of AGI, and there is no shared timetable that settles the question. On September 16, Legg said he assigned a 50% chance to “minimal” AGI by 2028. That is his forecast, not a scheduled release or an official Google prediction.
Legg’s warning: capability cannot outrun safety
Legg’s central message at launch was that progress in AI capabilities must not move faster than safety controls. He described the current period as one of rapidly advancing capabilities and argued that safety needs to keep pace.
That concern is reflected in the institute’s focus on reasoning transparency: understanding how advanced models reach their conclusions and whether that understanding can support safer deployment. The launch discussion also treats guardrails for systems that plan and complete tasks autonomously as a difficult technical problem, not a box that can simply be checked once.
The issue is not limited to laboratory performance. If systems become more capable across software, science, business and physical-world tasks, questions about who controls them, how their behavior is evaluated and how failures are contained become policy questions as well as engineering questions.
Economic policy depends on the scale of disruption
DMI’s economic-policy essay evaluates 11 household-facing interventions using literature reviews, surveys and 51 AI agent raters modeled on survey data from 51 economists. Rather than recommend one universal program, it links different responses to different scenarios:
| Disruption scenario described in the essay | Policy responses discussed |
| Mild disruption | Expanded unemployment insurance, a broader Earned Income Tax Credit and employer-led retraining |
| Moderate displacement and wage compression | A Negative Income Tax |
| Sustained labor-capital decoupling | A Universal Basic Capital backstop |
The essay reports that Universal Basic Capital received a composite score of 76.3 out of 100 for agency in the economist-panel analysis, the highest score in that category, but 33.1 out of 100 for feasibility. The contrast captures the problem the institute is trying to examine: a policy can give people more control in theory while remaining difficult to implement at scale.
The survey results point in a similar direction. Support reached 85% for publicly funded retraining, 72% for unemployment insurance and 54% for Universal Basic Capital. Those figures describe the essay’s survey findings; they do not predict which policy governments will adopt.
The launch puts that trade-off at the center of the AGI discussion. Technical progress, safety controls and economic preparation are being treated as connected problems—but the institute’s own policy work argues that no single intervention can answer all of them.
