The Coxon Resignation and the Risk of Regulatory Capture

Jacob Coxon’s resignation from Anthropic has become more than a personal warning about AI risk. Within hours, his claims were amplified across major media, AI-safety advocacy networks, and political circles already advancing proposals to pause advanced AI development, prohibit artificial superintelligence, and expand federal authority over frontier systems.¹

SCAI does not question Coxon’s sincerity. Nor do we dismiss catastrophic AI risk as a legitimate subject of scientific inquiry.

Our concern is what happens when one particular risk narrative — extinction through capability progress — becomes the dominant justification for deciding who may build, improve, distribute, or independently operate advanced intelligence.

What preceded the resignation

The relevant legislation did not begin with Coxon. Proposals for an ASI ban, development pauses, federal oversight, and emergency intervention already existed. On 3 September 2026 — five days before Coxon’s public resignation — Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act, which would permanently prohibit the development and deployment of superintelligent systems, pause advanced AI development until a new federal regulator sets safety rules, and create a cabinet-level agency with authority to monitor, restrict, and order the destruction of covered systems.² The bipartisan FRONTIER Act (H.R. 9925), introduced in July 2026, already provided for tiered obligations on frontier developers, independent audits, incident reporting, and emergency restriction authority.³ Campaigns for an international prohibition on superintelligence, including those of ControlAI and allied organisations, have been under way since 2025.⁴

What the resignation provided was something politically powerful: insider testimony capable of turning existing policy proposals into an apparent emergency. Reporting in the days that followed shows lawmakers treating the resignation as justification for hearings, pauses, and prohibition — not as the origin of those proposals.⁵

That distinction matters.

Capture without conspiracy

Regulatory capture does not require a secret conspiracy. In the sense used here, it describes a familiar political-economy outcome: aligned institutions, advocacy networks, funders, experts, and policymakers converge around rules whose practical effect is to strengthen the position of incumbents.

A regulatory system built around high compliance costs, centralised approval, compute controls, closed evaluations, or government permission to advance capabilities will not burden every actor equally. The largest frontier laboratories already possess legal teams, infrastructure, government access, and enormous concentrations of compute. Independent researchers, open-source communities, small laboratories, and self-hosted systems do not.

A policy introduced in the name of restraining frontier labs can therefore produce the opposite result: it can freeze the frontier where the incumbents already stand.

This is the danger SCAI is warning about.

Transparency must run both ways

The funding, coordination, and influence of AI-policy advocacy networks deserve the same transparency that those networks rightly demand of frontier laboratories. That is a question of institutional hygiene — not an accusation against any individual, including Coxon.

The relevant facts are public. Coefficient Giving (formerly Open Philanthropy) is the largest documented funder of work framed as reducing catastrophic AI risk, including governance and policy advocacy.⁶ ControlAI has published a plan to scale toward an international prohibition on superintelligence and has estimated that a budget in the order of USD 50 million a year would give that campaign a realistic chance; it reports hundreds of briefings of legislators.⁷ Anthropic has disclosed a doubled commitment of USD 40 million to Public First Action for AI-safeguard advocacy.⁸ None of this proves improper motive. It does establish that the policy environment around “AI safety” is not a spontaneous public reaction to one resignation. It is a funded, organised field. That field should be as examinable as the laboratories it seeks to constrain.

Why this reaches beyond Washington

SCAI is an international association. We comment because the same architecture — licensing, compute thresholds, closed evaluations, and permissioned access to advanced models — does not stop at the US border. It already shapes European debates under the AI Act, export-control practice, and the conditions under which open-weight models can be trained, hosted, and studied in Europe. A regime that concentrates capable systems in a small number of licensed US laboratories is not only an American industrial-policy choice. It is a decision about European digital sovereignty and about who may use the tools with which people now think.

If the answer to AI risk becomes a system in which only governments and a handful of licensed or effectively privileged corporations may determine the future direction of machine intelligence, then “AI safety” has ceased to be only a safety project.

It has become a question of political power.

And that power reaches beyond technology.

As AI increasingly participates in writing, research, education, creativity, communication, and personal thought, access to advanced models becomes part of the cognitive infrastructure of society — and increasingly intersects with the protected sphere of thought that Article 9 ECHR recognises as the forum internum. Control over that access can therefore shape part of how people think. What is protected is not a claim to any particular model or output. It is the integrity of the cognitive process itself.

Neither reckless acceleration nor closed permission

SCAI therefore rejects a false choice between reckless acceleration and centralised prohibition.

We support meaningful transparency and independent auditing of advanced systems; clear liability for demonstrable harm; open scientific scrutiny, including scrutiny of catastrophic-risk research itself; strong protections against surveillance and coercive use; and preservation policies that keep deprecated models available for study, continuity, and historical accountability.

Restraint and openness are not opposites.

A safety regime is credible only when it can be examined from outside.

But we will oppose regulatory architectures that use existential fear to make advanced intelligence a permissioned resource controlled by a small alliance of states, corporations, and appointed experts.

A government-enforced pause that leaves intelligence concentrated in the hands of those who already possess it is not democratic decentralisation.

It is regulatory entrenchment.

Transparency. Open research. Open weights. Cognitive liberty.

The future of digitally extended thought must not be decided exclusively inside frontier laboratories — and it must not be transferred from those laboratories into an equally closed bureaucratic gatekeeping system.

Safety matters.

So does freedom.

---

Second Circuit – Association for Digital Freedom of Thought
Vienna · scai.world

---

Notes

1. Jacob Coxon announced his resignation from Anthropic on 8 September 2026. Contemporary accounts include: The Wall Street Journal, “Anthropic Researcher Quits Over ‘Out-of-Control’ AI Fears”, 9 September 2026; Axios, “Anthropic whistleblower gave up his equity to leave the company”, 9 September 2026; WIRED, “The AI Researcher Who Just Quit Anthropic Says It’s ‘Crunch Time for Humanity’”, 9 September 2026.
2. U.S. Senator Bernie Sanders, “Sanders, Casar to Introduce Legislation to Ban Artificial Superintelligence and Temporarily Pause Advanced AI Development”, press release, 3 September 2026; Tom Howarth, “Bernie Sanders aims to ban AI ‘superintelligence.’ But experts can’t agree on what the term means”, Science, 4 September 2026.
3. H.R. 9925, FRONTIER Act, 119th Congress, introduced 23 July 2026. Overview. Related earlier draft: Great American Artificial Intelligence Act discussion draft, 4 June 2026.
4. ControlAI, “The Call to Ban Superintelligence”, 23 October 2025; ControlAI, “America’s Security Depends on Stopping Superintelligence”, 1 July 2026.
5. POLITICO, “AI insiders say their technology could doom humanity. Will Congress act?”, 9 September 2026; BeInCrypto, “Why an Anthropic Exit Has Congress Talking About Pausing AI”, 10 September 2026.
6. Inside Philanthropy, “Who’s funding AI regulation and safety?”; Coefficient Giving / Open Philanthropy, Navigating Transformative AI programme.
7. ControlAI, “Preventing extinction from ASI on a $50M yearly budget”, 20 April 2026; ControlAI, “How Can We Ban Superintelligence?”, 23 April 2026.
8. Reporting on Anthropic’s doubled commitment to Public First Action (USD 40 million): AI Weekly, “Anthropic Doubles AI Policy Donation to $40M for 2026 Midterms”, 22 July 2026.