Version 3

The AI safety governance debate has extended into questions of whether stated ethical commitments hold under financial and political pressure. TurnTrout, a researcher who left Google DeepMind, documents the organization signing a classified Pentagon deal permitting 'any lawful government purpose' without restrictions on autonomous weapons or mass surveillance — while CEO Demis Hassabis publicly claimed 'nothing's changed about our principles' [8]. This contrasts directly with Anthropic ending Pentagon negotiations in February 2026 over the same issues [7]. OpenAI separately published advocacy for 'reverse federalism,' arguing California, New York, and Illinois AI safety laws should converge into a de facto national standard ahead of formal federal legislation, with the Trump administration targeting early August 2026 for a federal model-testing framework [9]. These developments sit alongside an ongoing split between Plan A's cooperative US-China governance approach and the domestic competitive controls — export restrictions and a possible open-weight model ban — favored by the current administration [2][6].

TurnTrout's account introduces empirical evidence that AI safety commitments can be abandoned under institutional pressure, which bears directly on whether voluntary industry governance or formal binding mechanisms are viable. If OpenAI's reverse federalism strategy takes hold, three states may set the practical terms of frontier AI safety regulation before any federal framework exists.

The central debate in AI safety governance is between advocates of international cooperative controls and those pursuing domestic competitive strategies. The AI Futures Project's Plan A proposes that the US and China jointly control chip supply, audit data centers, and share research to slow superintelligent AI development [1]. Zvi Mowshowitz argues Plan A deserves serious engagement rather than dismissal, with the key dividing question being whether superintelligence arrives soon enough to justify Plan A's costs [2]. Vitalik Buterin defends Plan A against critics who call it naive, arguing they apply coordination skepticism to a cooperative pause but not to the assumption that an AI transition will simply go well by default — where humanity's hard power drops to zero if AIs outperform humans at every task [2]. The Trump administration has not engaged with the cooperative frame, governing AI through competitiveness: a January 2025 policy framework, export controls on model weights effective January 2026, preemption of state AI regulations, and a reported White House discussion of an executive order to ban open-weight frontier models outright [3][4][5][6].

The question of whether AI safety commitments hold under institutional pressure has taken concrete form. Anthropic ended Pentagon contract negotiations in February 2026 after the Defense Department demanded blanket 'anything lawful' usage rights, with Dario Amodei citing irreconcilable conflicts with Anthropic's redlines on mass surveillance and autonomous weapons targeting [7]. Google DeepMind took the opposite path: a former researcher writing as TurnTrout published a detailed account of leaving the organization after Google signed a classified Pentagon deal permitting 'any lawful government purpose' without binding restrictions on autonomous weapons or mass surveillance [8]. TurnTrout documents CEO Demis Hassabis removing specific prohibitions from Google's 2018 AI principles while publicly claiming 'nothing's changed about our principles,' and reports that the IASEAI publicly promised a member poll supporting Anthropic's stance but silently cancelled it without explanation [8]. The piece frames the divergence plainly: 'When profit and pressure met ethical commitment at Google DeepMind, pressure won and pledges lost. When profit and pressure met ethical commitment at Anthropic, ethics won' [8].

A separate domestic debate concerns open-weight frontier models. Nathan Lambert reports White House discussions of an executive order that could ban open-weight models above current frontier capability levels, and accuses Anthropic's campaign against Chinese model distillation of constituting regulatory capture — noting Anthropic would gain substantial commercial benefit if the targeted Chinese model makers were banned [6]. Lambert disputes the security rationale by pointing to Anthropic's Mythos model being accessed through unauthorized Discord channels during private beta, suggesting APIs are not meaningfully more secure than open weights [6]. OpenAI, for its part, published advocacy for 'reverse federalism': state-level AI safety laws in California, New York, and Illinois — requiring documented safety frameworks, public risk assessments, incident reporting, and independent audits — positioned as converging toward a de facto national standard ahead of federal legislation, with the Trump administration separately targeting early August 2026 for a cyber-evaluation framework [9]. A parallel argument on the Alignment Forum holds that political will, not research, is the main AI safety bottleneck, with a 3.6:1 researcher-to-advocate ratio in the US AI safety field and AI companies securing seven times as many European Commission meetings as civil society in 2023 [10].