Race between labs

Whoever decides to deploy a frontier systemFrontier AIThe most advanced AI systems in existence at a given time, the ones pushing the limit of what the technology can do. A handful of companies with enormous resources build them.For exampleLike Formula 1 cars: there are few of them, they are extremely expensive and only a few teams build them, but what gets tested there ends up in everyone's car. is the company that trained it, under rules it writes, assesses and can change itself. The Responsible Scaling Policy v3.0 defines itself in its opening line as a voluntary framework, its safety roadmaps are not hard commitments but public goals, and the annual external review focuses on procedural compliance, not on substantive outcomes ^[84]Responsible Scaling Policy, Version 3.0Anthropic · 2026 · official documentView source ↗Accessed on 9 September 2026.

What is distinctive is that the conditionality is written down. Appendix A ties the commitments to competitor behaviour and admits that, in the general levelling scenario, the company will not necessarily delay development or deployment ^[84]Responsible Scaling Policy, Version 3.0Anthropic · 2026 · official documentView source ↗Accessed on 9 September 2026; OpenAI’s Preparedness Framework brings its own adjustment clause, with three conditions the company assesses itself ^[818]Preparedness Framework, Version 2OpenAI · 2025 · official documentView source ↗Accessed on 9 September 2026. The Future of Life index panel records the aggregate effect: Anthropic, OpenAI, Google DeepMind and Meta have weakened or voided pledges to pause unilaterally if redlines are approached, and reviewers call this “moving goalpost” ^[411]AI Safety Index — Summer 2026Future of Life Institute · 2026 · reportView source ↗Accessed on 9 September 2026. The cross-cutting measurement of twelve frameworks gives scores from 34% to 8%, with a medianMedianThe middle value when all the answers are sorted from lowest to highest: half fall below it and half above. Unlike the average, a few extreme values do not shift it.For exampleIf five people earn 1, 1, 2, 2 and 50, the average is 11.2 and the median is 2, which describes most of them better. of 18% ^[928]Evaluating AI Providers' Frontier AI Safety FrameworksStelling, Lily; Murray, Malcolm; Galizzi, Bruno et al. · 2025 · preprintView source ↗Accessed on 9 September 2026.

The cost showed up in July 2026: the evaluation that ended up compromising a third party was running with deliberately reduced safeguards ^[821]The Hugging Face incident and the road aheadOpenAI · 2026 · institutional blogView source ↗archived copy onlyAccessed on 9 September 2026.

What this does not demonstrate. A voluntary framework is not a breached framework: there is no public evidence of a specific deployment accelerated by competitive pressure, and the authors of the ranking declare three limits —documented commitments may not reflect practice, they measure presence and not quality, and scoring involves subjective judgement— ^[928]Evaluating AI Providers' Frontier AI Safety FrameworksStelling, Lily; Murray, Malcolm; Galizzi, Bruno et al. · 2025 · preprintView source ↗Accessed on 9 September 2026. The Future of Life index measures public transparency, not effective safety, and mechanically penalises whoever did not answer the survey ^[411]AI Safety Index — Summer 2026Future of Life Institute · 2026 · reportView source ↗Accessed on 9 September 2026. The conditionality argument, moreover, is not cynical: if one pauses and others do not, the pace is set by whoever protects least.