Google DeepMind

Stated Theory of Change

Google DeepMind's stated theory is a three-part argument:

1. AGI will be transformative and necessary. Hassabis frames AGI as "the cavalry" -- the technology needed to solve climate change, cure diseases, and enable "radical abundance." Without it, he would be "very worried about society." This makes building AGI a moral imperative, not just a commercial one.

2. Safety is best achieved from inside the frontier lab. By embedding safety researchers directly in the lab building AGI, they have maximum influence over deployment decisions. Dafoe's theory: "advising influential decision makers" who have "the right character." Shah's argument: alignment is commercially motivated because misalignment is currently the bottleneck to useful products.

3. A framework of escalating commitments will catch dangerous capabilities before they cause harm. The Frontier Safety Framework defines Critical Capability Levels with thresholds that trigger security and deployment mitigations. The 145-page AGI safety paper outlines defense-in-depth across four risk areas.

The implicit assumption underlying all three parts: the people running GDM (Hassabis, Legg) genuinely care about safety and will use their influence to make the right calls when commercial and safety interests conflict.

Revealed Theory of Change

Actions tell a different story than statements. The revealed theory appears to be:

"Move as fast as corporate constraints allow, publish safety research to maintain credibility, and trust that alignment-as-product-quality will keep pace with capability scaling."

Evidence for the gap between stated and revealed theories:

• Weapons pledge dropped (Feb 2025): Hassabis co-authored the blog post removing Google's 2018 commitment against weapons and surveillance AI. Two years earlier, he warned about "not moving fast and breaking things." The geopolitical justification ("world's become a much more dangerous place") applies equally to all labs -- it doesn't explain why GDM specifically reversed course.

• Model card delays/omissions: Gemini 2.5 Pro released without the safety documentation GDM committed to at the White House, G7, and Seoul. This was not an oversight -- it was a speed-over-transparency decision.

• Safety data withheld: Gemini 3's safety report systematically hides manipulation efficacy, propensity rates, and detailed evaluation results. If the numbers were good, they would be published.

• Project Nimbus signed despite internal warnings: Google's own consultants recommended withholding AI from the Israeli military. Google signed a potentially 23-year contract anyway. Protesters were fired. Internal discussion was restricted.

• Independence bid rejected and not pursued further: After Alphabet rejected DeepMind's bid for independent legal structure in 2021, Hassabis appears to have accepted the outcome rather than escalating publicly.

The revealed theory of change is more honest than it sounds. Many GDM researchers likely believe that being inside Google -- with access to compute, data, and deployment infrastructure -- gives them more total safety impact than they would have independently, even if they lose individual battles (weapons pledge, model card timing). This is Allan Dafoe's explicit theory. The question is whether the battles being lost are ones that matter.

Key Assumptions

Assumption 1: Hassabis and Legg will have the power to override commercial pressure on safety.

• Evidence for: Hassabis is CEO with Nobel Prize credibility. Legg leads AGI Safety Council. Both signed CAIS statement.
• Evidence against: Weapons pledge dropped. Independence bid rejected. Pichai has ultimate authority. Hassabis calls Pichai daily -- who is influencing whom?
• Testable: Would become clear if GDM ever delayed or halted a release for safety reasons against commercial pressure. This has not happened.
• If wrong: Safety decisions are effectively Alphabet board decisions, made by people optimizing for shareholder returns.

Assumption 2: Alignment research at a frontier lab has more impact than external safety work.

• Evidence for: Shah's point about needing access to frontier models for safety research. Dafoe's "in the room where it happens" argument.
• Evidence against: The Gebru/Mitchell precedent shows Google fires researchers whose findings embarrass products. The "Right to Warn" letter shows employees feel unable to speak freely.
• Testable: Does GDM's safety research actually change deployment decisions? So far, no public example exists.
• If wrong: Safety researchers at GDM are providing legitimacy to unsafe practices rather than preventing them.

Assumption 3: The FSF will function as intended when CCLs are triggered.

• Evidence for: The framework exists and is updated (3 versions). Uniquely addresses deceptive alignment. Applies to all of Google.
• Evidence against: Uses "aim" not "commit." 6x compute gap is the widest among top labs. No concrete mitigation plans. Never been tested.
• Testable: Will be tested when a model actually triggers a CCL. Has not happened yet.
• If wrong: The FSF is a public relations document, not a safety framework.

Assumption 4: Alignment research will scale with capability scaling (the "commercial alignment" thesis).

• Evidence for: Currently true -- misalignment causes product failures, creating commercial incentive for alignment.
• Evidence against: This only holds while models are weak enough that misalignment causes visible product failures, not invisible catastrophes. As models become more capable, misalignment may become harder to detect but more dangerous.
• If wrong: The commercial case for alignment evaporates precisely when alignment matters most.

Strengths

Talent depth. GDM has arguably the deepest bench of AI safety talent at any frontier lab: Shah, Dafoe, Dragan, Orr, Nanda, Gabriel, plus Legg as executive sponsor. Multiple of these people (Shah, Dafoe) are known for intellectual honesty and genuinely seem to be trying to do good work.

Scientific credibility. AlphaFold is the strongest real-world AI safety argument: proof that AI can produce massive public benefit. The Nobel Prize gives Hassabis credibility that other frontier lab CEOs lack. GDM's research culture, while compressed by the merger, still produces fundamental science.

Infrastructure advantage. Vertical integration of custom chips (TPUs), data centers, and models gives GDM structural advantages no competitor can replicate quickly. This means GDM can theoretically afford to invest in safety without existential competitive pressure.

Framework applies to all of Google. Unlike OpenAI's framework (which doesn't apply to Microsoft), GDM's FSF covers the entire Google ecosystem. This is a genuine structural advantage.

Deceptive alignment as explicit risk class. GDM is the only frontier lab that explicitly identifies deceptive alignment in its safety framework. This suggests at least some of the right threat models are informing the framework design.

Weaknesses and Risks

No structural independence from Alphabet. This is the fundamental problem. Every other weakness flows from it. Hassabis cannot make a safety decision that costs Alphabet money without Pichai's approval. The weapons pledge reversal demonstrates that when commercial interests and safety commitments conflict, commercial interests win.

Pattern of broken commitments. The 2018 weapons pledge lasted 7 years before being dropped. Model card commitments were violated within a year of being made. The Seoul Commitments were not honored for Gemini 2.5 Pro. Each broken commitment erodes the credibility of all remaining commitments.

Safety opacity. GDM systematically withholds safety data: manipulation efficacy numbers, propensity rates, detailed evaluation results. The FLI Safety Index specifically recommended more transparent and independent evaluations. The Gemini 3 safety report "hides the football" repeatedly.

Talent hemorrhage. 142 departures post-merger. All transformer paper authors gone. Noncompete enforcement creates resentment. The researchers who produced AlphaFold worked under conditions (research freedom, minimal commercial pressure) that no longer exist.

"Aim" not "commit." The FSF's aspirational language means no commitment is enforceable. Combined with no independent governance, this means safety is a preference, not a constraint.

Suppression of internal dissent. Gebru/Mitchell firings. 50+ employees fired for protesting Nimbus. Political discussion restricted on internal forums. "Right to Warn" letter went unanswered. This environment is hostile to the kind of internal challenge that safety requires.

Cross-References

vs. Anthropic: Anthropic was founded by people who left Google over exactly the concerns that define GDM's weaknesses: structural independence, safety culture, deployment decisions. Anthropic has the LTBT, independent board, RSP with "commit" language, and safety as core identity. GDM has none of these structural features. But GDM has compute advantages Anthropic cannot match and scientific credibility (AlphaFold, Nobel) that Anthropic doesn't have.

vs. OpenAI: Both are part of larger corporate structures (Alphabet and Microsoft respectively). OpenAI's Preparedness Framework checks every 2x compute (vs GDM's 6x). OpenAI has been more publicly messy (board drama, departures), which paradoxically provides more visibility into internal tensions. GDM's smooth exterior may hide more than it reveals.

Allan Dafoe connection to GovAI: Dafoe founded GovAI, then left for GDM. This represents a theory-of-change bet: that insider influence at a frontier lab matters more than independent governance research. If GDM's safety culture degrades, this bet looks worse in hindsight.

Google's $2B Anthropic investment: This hedging strategy means Google benefits if the safety-first approach succeeds. It also means Google is funding a competitor that makes GDM look less safety-conscious by comparison.

What Would Change This Assessment

Positive updates (would increase confidence in GDM's safety approach):

• GDM publicly delays or halts a model release for safety reasons, against commercial pressure
• FSF language changes from "aim" to "commit" with enforceable mechanisms
• Independent safety board with veto power established
• Detailed, unredacted safety reports published for all major releases
• Hassabis advocates publicly for regulation, even if it disadvantages Google

Negative updates (would decrease confidence):

• Another major commitment broken (Seoul, UK AISI partnership, etc.)
• Senior safety researcher departs publicly citing safety concerns
• GDM deploys a model that triggers a CCL but proceeds anyway
• Further suppression of internal safety dissent
• Alphabet pressures GDM to reduce safety team or budget

Self-Critique

Weakest claim: The assertion that GDM's safety commitments are primarily performative. It is genuinely possible that the FSF, AGI safety paper, and safety research program represent a serious institutional effort that simply can't be fully public due to competitive dynamics. Rohin Shah and Allan Dafoe are credible, thoughtful people who appear to genuinely believe their work matters. If they are right that insider influence is the highest-impact strategy, then this analysis overstates the importance of structural independence.

Potential bias: This analysis may be overly influenced by the weapons pledge reversal and Project Nimbus, which are vivid, concrete examples that anchor criticism. GDM's day-to-day safety work -- which is less visible -- may be more substantial than these high-profile failures suggest.

What I would most want to check: Internal resource allocation data. If GDM spends 10%+ of its budget on safety research and evaluation, the picture looks very different than if safety is 1%. Without financial transparency, this cannot be assessed.

What a thoughtful defender would say: "GDM has the deepest safety research team at any frontier lab, produces genuinely important safety research (debate, interpretability, dangerous capability evaluations), and the FSF -- even with 'aim' language -- represents more systematic safety thinking than most labs have. The weapons pledge and Project Nimbus are Google decisions, not GDM decisions. Hassabis and Legg are doing the best they can inside a corporate structure they don't control, and their presence at the frontier makes the overall trajectory safer than if a less safety-conscious team were building these systems."

What a thoughtful critic would say: "GDM's safety work provides legitimacy to a corporation that has broken every safety commitment it has made. Hassabis co-authored the blog post dropping the weapons pledge. The FSF uses 'aim' not 'commit.' Safety data is systematically withheld. The most important safety decision GDM ever had to make -- whether to accept corporate capture or insist on independence -- was made in 2021 when they accepted Alphabet's rejection of independent legal structure. Everything since then is operating within constraints that make genuine safety governance impossible."