When Safety Becomes a Competitive Disadvantage: The Structural Dangers of the Frontier AI Race

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Frontier artificial intelligence is commonly described as a race. The central risk is an incentive structure in which additional evaluation, delayed deployment, disclosure, or binding oversight can impose concentrated costs on a cautious developer while distributing much of the safety benefit to others. This paper identifies seven conditions that intensify that hazard: a large relative prize, a speed premium, assurance-time substitution, rival uncertainty, externalised harm, weak reciprocity, and difficult-to-reverse capability diffusion. It examines published frameworks, experimental research, incident reports, and the September 2026 disagreement over coordinated pacing. That disagreement exposes competing accounts of whether customer demand and liability sufficiently reward safety or whether reciprocal constraints are also required. Operational disclosures motivate a further hypothesis: development can generate assurance obligations faster than organisations can discharge them. Neither those disclosures nor public endorsements establish that competition caused a particular incident or that a coordinated slowdown has occurred. The paper retains counterevidence, including reported costly pauses and methods that improve evaluation efficiency. It extends a qualitative Race Hazard Audit with backlog, evidence-freshness, oversight-latency and evaluator-independence questions. The proposed governance standard is to make necessary restraint reciprocal, verifiable, competitively survivable and enforceable, while preserving beneficial competition. The framework remains a research proposal rather than a validated risk score or causal estimate. Version 0.2 — 19 September 2026. Revised working paper with an evidence cutoff of 19 September 2026. Retains the original seven-condition framework and six failure pathways; adds the pacing debate, research and evaluation incidents, assurance-backlog analysis, evaluator funding and authority, policy updates, and tests capable of challenging the thesis. This is a conceptual working paper, not peer-reviewed empirical validation. AI-assisted research and drafting are disclosed in the manuscript.