Regulatory capture

There is no need to assume bad faith: it is enough to look at who writes the rule, who assesses it and who can change it. Today they are the same actor. The cross-cutting evaluation of twelve frontier safety 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%, and three findings that recur: most providers lack independent oversight of their safety decisions, commitments were sometimes weakened without clear justification, and providers retain discretion at the key decision points ^[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 of July 2026 records the 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.

On the international side, follow-up does not exist. The New Delhi commitments of February 2026 are two in number and voluntary, and neither addresses capability thresholds, catastrophic risk or concentration of power ^[854]Championing Inclusive and Multilingual AI for the Global South, India Unveils New Delhi Frontier AI CommitmentsMinistry of Electronics and Information Technology, Government of India · 2026 · official documentView source ↗archived copy onlyAccessed on 9 September 2026; there is no accountability framework to track deliverables from one summit to the next, no central secretariat and no coordinating body ^[1021]Multistakeholder Promises and Power Gaps in Global AI SummitsHickok, Elonnai; Mohan, Shashank; Pielemeier, Jason et al. · 2026 · pressView source ↗Accessed on 9 September 2026.

What this does not demonstrate. “Self-assessed” is not the same as “captured”: the regulated party writing the rule does not prove that the rule serves its interests against the public interest. The European case cuts both ways. Regulation (EU) 2026/1744 deferred the Annex III high-risk obligations from August 2026 to December 2027, but in the same text it added two prohibited practices —non-consensual intimate images and child sexual abuse material— applicable from December 2026 ^[25]Digital Omnibus on AI — AI Act ExplorerEU Artificial Intelligence Act (proyecto de referencia) · 2026 · webView source ↗Accessed on 9 September 2026. And the strong version of the thesis, that companies push regulation in order to shut competitors out, circulates as a hypothesis in technical forums without any measurement at all. In the United States the observable outcome is simpler: the two benchmarkBenchmarkA standardised test for measuring and comparing what different AI models can do. When models solve almost all of it, it is called “saturated” and stops telling them apart.For exampleLike a university entrance exam: everyone answers the same questions, so the scores can be compared. federal bills remain at the first stage with no committee action ^[469]H.R. 9917: AI Kill Switch ActGovTrack.us · 2026 · webView source ↗Accessed on 9 September 2026 ^[468]S. 1792: AI Whistleblower Protection ActGovTrack.us · 2026 · webView source ↗Accessed on 9 September 2026.