Google DeepMind Launches the DeepMind Institute — Hassabis Proposes a US-Led Standards Body With 30-Day Pre-Release Model Evaluation
Google and Google DeepMind researchers launched the DeepMind Institute on Wednesday, September 16, 2026 — a public forum intended to "surface differing views between Google, Google DeepMind, and the broader global research community around AGI." Per TechCrunch (Sept 17), its directors are DeepMind co-founder Shane Legg (also managing editor), Google executive James Manyika, and Google DeepMind chair Demis Hassabis. Google's own framing: "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."
The inaugural collection is four essays — AGI economic policy, preserving human-readable model reasoning, human flourishing principles, and a frontier-model evaluation framework. Two carry concrete governance proposals:
- Hassabis proposes a US-led standards body to evaluate the most advanced AI models. Developer companies would voluntarily submit models for testing up to 30 days before release; at first the body would design evaluations in consultation with AI companies, but it would eventually develop independent, undisclosed "held-out" tests so labs can't tailor models to known evaluations.
- DeepMind safety researchers Rohin Shah and Anca Dragan argue the loss of verifiable reasoning is not inevitable: regulators could limit "opaque sequential depth" — the amount of sequential computation a model can perform without a readable reasoning trace — or require developers to demonstrate that less transparent systems remain equally monitorable.
Why it matters: Google was the conspicuous holdout from concrete pacing proposals while Amodei's framework, OpenAI's disclosure framework (see OpenAI Ships Misalignment Disclosure Framework and Six Incident Reports — Weeks Ahead of Schedule), and Anthropic's metrics (see Anthropic's R&D Automation Index: Claude Now "Leads" 26% of Its Own R&D — First Public Pace Metrics From a Frontier Lab) accumulated. This is Google's entry into mechanism design, and its 30-day pre-release evaluation window converges with the audit-before-release direction of the Thune-Cruz-Klobuchar Senate bill and the embedded-evaluator commitments both OpenAI and Anthropic have made. The debate TechCrunch describes has shifted "from broad statements of concern toward concrete proposals for disclosure, outside scrutiny, and, if safeguards fall behind, coordinated slowdowns" — and the industry's three biggest labs have now each put a specific mechanism on the table.