Google DeepMind

Theory of Change

Google DeepMind's theory of change is that building AGI -- which Hassabis defines as AI capable of genuine scientific hypothesis generation, not just economically useful labor -- is the fastest path to solving humanity's most pressing problems. In Hassabis's words: "I'd be very worried about society today if I didn't know that something as transformative as AI was coming down the line. I firmly believe that. It's almost like the cavalry."

The safety theory of change is that (a) alignment research is commercially motivated because misalignment is currently the bottleneck to useful products, (b) embedding safety researchers inside the lab building AGI gives them maximum influence over deployment decisions, and (c) the Frontier Safety Framework creates thresholds that trigger mitigations before dangerous capabilities emerge. The 145-page "Approach to Technical AGI Safety and Security" paper (April 2025) outlines defense-in-depth: amplified oversight, robust training, interpretability, and system-level security across four risk areas (misuse, misalignment, mistakes, structural risks).

Co-founder Shane Legg, who has been publicly concerned about AI existential risk since 2011, leads the AGI Safety Council. Allan Dafoe, founding director of GovAI, left academia to join GDM because he believes "Demis and Shane are the right leaders" and his impact is maximized by advising influential decision-makers from inside the lab. Rohin Shah frames the core argument: "Alignment work is sort of the bottleneck to actually getting useful products out into the world."

What They Do

Products. Gemini model family (3 Pro released Nov 2025) now matches or exceeds state-of-the-art on most benchmarks. 650+ million monthly active users. ChatGPT still holds 64.5% market share vs Gemini's 21.5%. Google embeds AI across Search (8.5B queries/day), Workspace (1B+ users), Android (2B devices), and Cloud ($70B annual run rate).

Science. AlphaFold predicted 200M+ protein structures, freely available, used by 3M+ researchers in 190 countries. Nobel Prize Chemistry 2024 for Hassabis and Jumper. Isomorphic Labs (drug discovery spinoff, raised $600M) entering human clinical trials with AI-designed drugs. AlphaEvolve solved open mathematical conjectures. Weather Lab outperformed physics-based models for cyclone forecasting.

Safety Research. Frontier Safety Framework published in three versions (June 2024 to Jan 2026) with Critical Capability Levels in autonomy, biosecurity, cybersecurity, ML R&D, and manipulation. Published 145-page AGI safety paper. Gemma Scope (400+ sparse autoencoders for interpretability). Scalable oversight research. Doubly-efficient debate protocol. MONA for reward hacking mitigation. Uniquely among frontier labs, GDM explicitly identifies deceptive alignment as a risk class.

Commitments Record. Signed CAIS extinction risk statement, Seoul Safety Commitments. Released Gemini 2.5 Pro without model card, violating White House, G7, and Seoul commitments (60 UK MPs accused Google of "breach of trust"). February 2025: dropped 2018 AI Principles pledge against weapons and surveillance, co-authored by Hassabis himself. Gemini 3 safety report withheld key safety numbers including manipulation efficacy data. Earlier precedent: in 2015-2017, DeepMind's Streams app accessed 1.6 million NHS patient records without adequate consent, leading to ICO findings of non-compliance. The data was subsequently transferred to Google Health, contradicting privacy assurances -- establishing a pattern of governance failures around sensitive data that predates the current safety structures.

Key People

Demis Hassabis (CEO, co-founder): Neuroscience PhD, chess prodigy, Nobel Laureate. Identifies as "a scientist first." Sets a higher bar for AGI than competitors ("Could an AI have come up with general relativity?"). Co-authored blog post dropping Google's weapons pledge. Calls Sundar Pichai daily.

Shane Legg (co-founder, Chief AGI Scientist): PhD in machine super intelligence. Concerned about AI x-risk since 2011. Predicted 50% AGI by 2028. Leads AGI Safety Council. The most x-risk-concerned co-founder at any major frontier lab.

Notable departures. Geoffrey Hinton left Google May 2023 to warn about AI risks: "I want to talk about AI safety issues without having to worry about how it interacts with Google's business." Won Nobel Physics 2024. Mustafa Suleyman (co-founder) departed 2019 amid bullying complaints, now CEO of Microsoft AI. 142 departures after the Brain/DeepMind merger. All 8 authors of the transformer paper ("Attention Is All You Need") have left. Noncompete enforcement (6-12 months, UK) is controversial. Counterpoint: 20% of 2025 AI hires were returning employees.

Safety team. ~30-50 people in core AGI alignment (Anca Dragan leading, Rohin Shah, Allan Dafoe, Dave Orr). Growing ~37-39% per year. Total GDM headcount: 6,000-7,600. No published safety-to-total ratio.

Money and Incentives

Corporate structure. Wholly-owned subsidiary of Alphabet Inc. (market cap ~$2.4T). No independent financials, no independent legal structure. DeepMind's 2021 bid for independent legal structure was rejected by Alphabet. All decisions ultimately answer to Alphabet's board and shareholders.

Scale of investment. Alphabet planned 2026 capex: $175-185B, mostly for AI compute. Pre-merger DeepMind (2020): GBP 826M revenue, all from internal Google payments. Google Cloud: $70B annual run rate, $240B backlog. Google designs its own TPU chips (Ironwood: 4,614 TFLOPS per chip), giving GDM a structural hardware advantage no competitor can replicate.

Revenue model. AI is embedded across Google's existing products, not a standalone business. This means Google can afford to move cautiously -- but competitive dynamics with OpenAI drive urgency regardless.

Military contracts. Project Nimbus: $1.2B joint contract with Amazon providing cloud/AI to Israeli government and military. Google's own internal report acknowledged "Google Cloud Services could be used for, or linked to, the facilitation of human rights violations." Third-party consultant recommended withholding AI tools. Google signed anyway. Contract may extend up to 23 years. Google fired 50+ employees who protested. At least 5 UK DeepMind staff resigned. Google restricted political discussion on internal forums.

Weapons pledge reversal. In February 2025, Google removed its 2018 AI Principles pledge against weapons and surveillance. The blog post was co-authored by Hassabis. Justified by "global competition." Human Rights Watch: "Google announces willingness to develop AI weapons." Google employees' 2018 petition (thousands signed, dozens resigned) that originally created the pledge has been effectively overridden.

Hedging. Google invested $2B in Anthropic -- a company founded by people who left over safety concerns, now a direct competitor. This hedges Google's bet: if GDM's approach fails, Google still has a financial interest in the safety-focused alternative.

No external funding. Zero Coefficient Giving/Open Phil grants (expected for a Google subsidiary). The separate "DeepMind Foundation" (EIN 822828614) has ~$856K assets and makes $500-$1000 grants in medical research -- irrelevant to GDM operations.

What Others Say

FLI AI Safety Index (Winter 2025): Grade C (2.08/4.0), third behind Anthropic (C+) and OpenAI (C+). Recommendations: strengthen risk-assessment rigor, make governance structures more actionable, reduce lobbying against state-level AI safety regulations.

Zvi Mowshowitz on the Frontier Safety Framework: "This document is weak and unambitious. It is disappointing relative to my expectations." Key weaknesses: the word "commit" never appears; checks only every 6x compute (widest gap of top 3 labs); mitigation plans are empty; misalignment addressed as "future work."

Zvi on Gemini 3: "The model is seriously misaligned in many ways... prone to hallucinations, crafting narratives, glazing." Safety report "repeatedly hides the football." Dan Hendrycks: "On safety -- jailbreaks, bioweapons assistance, overconfidence, deception, agentic harm -- Gemini is worse than GPT, Claude, and Grok."

Semafor: "Hassabis paradox: warning about AI race while leading the race."

Stuart Russell (FLI panel): "AI CEOs claim they know how to build superhuman AI, yet none can show how they'll prevent us from losing control... I'm looking for proof that they can reduce the annual risk of control loss to one in a hundred million. Instead, they admit the risk could be one in ten, one in five, even one in three."

Rohin Shah (GDM insider): "I'm not really sold on [doom narratives]... I expect AI systems will become more powerful relatively continuously... I would be generally in favour of the entire world slowing down on AI progress." On alignment difficulty: "Nobody has very compelling arguments that AI will be dangerous by default, or that it will be safe by default. We just don't know."

Allan Dafoe (GDM insider): Technology is "unstoppable" in the sense that competitive dynamics force adoption, "but the FORM it takes is shapeable." His theory of impact: "advising influential decision makers" like Hassabis and Legg who have "the right character."

What's Absent

No independent safety governance. Unlike Anthropic (LTBT with board veto) or even pre-restructuring OpenAI (nonprofit oversight), GDM has no structural mechanism for safety to override commercial pressure. All governance is internal to Alphabet. When asked what happens if DeepMind disagrees with Alphabet on safety, COO Lila Ibrahim said: "We haven't had that issue yet."

Secret ethics board. The board created as a condition of the 2014 acquisition has never been publicly disclosed. Hassabis says it's "all confidential." Nobody knows who was on it, whether it still exists, or what it did.

No financial transparency. As an Alphabet subsidiary, GDM's budget, safety spending, and resource allocation are all opaque. No way to verify claims about safety investment.

No published safety-to-total headcount ratio. Estimates suggest 30-50 people in core alignment out of 6,000-7,600 total. No public commitment to maintain or grow this ratio.

No safety-driven non-deployment decision. GDM has never publicly declined to release a model for safety reasons. The FSF's "aim" language (not "commit") makes halt commitments unenforceable.

Missing safety data. Gemini 3 safety report withheld manipulation efficacy numbers, propensity rates, and detailed external evaluation results. Pattern of selective disclosure.

No policy for disagreements with Alphabet. No escalation mechanism, no safety veto, no dead man's switch for when commercial and safety interests conflict.