Anthropic chief executive Dario Amodei published an essay this month, “We Must Pace the Frontier,” calling for frontier AI development to slow down through a mix of voluntary commitments and regulation. Preston Byrne, a lawyer who has spent roughly the last 18 months representing clients against internet censorship regulators in the UK and Europe, argues on his personal blog that the plan, whatever its intentions, would rebuild that same enforcement apparatus around AI companies instead of social platforms.
Byrne lays out Amodei’s proposal fairly before attacking it. It rests on three parts: “Embedded Evaluators,” third parties such as METR given employee-like access to verify a lab’s safety practices and training pipelines; “Democratic Coordination,” in which frontier labs in allied countries agree to common safety standards, which Amodei says will need government backing because such coordination between competitors is “legally challenging”; and “Global Coordination,” an attempt to bring even authoritarian governments into a compliance framework. Amodei frames the slowdown as starting voluntary, but he is explicit that regulation is the mechanism that would eventually bind everyone, including labs unwilling to cooperate on their own.
That is the crux of Byrne’s objection. A rule written to apply only if a company refuses to volunteer is not voluntary. It is a legal obligation backed by the state, and Byrne treats it as such rather than accepting the softer framing in Amodei’s essay.
His skepticism is structural, not abstract. Byrne points to the Global Alliance for Responsible Media, a now-disbanded industry group that the House Judiciary Committee has described as part of a “censorship-industrial complex,” as the precedent for what “embedded evaluators” become in practice: private bodies with outsized influence over what a platform, or in this case a model, is allowed to do. He argues the “Democratic Coordination” plank creates two separate legal problems, an antitrust exposure for Anthropic and OpenAI if they align policies as the two largest labs, and a First Amendment problem in the United States if coordination extends to limiting how quickly software itself can be built, since he cites the line of cases including Bernstein v. United States and Smith v. California, plus the legislative history behind Section 230, as protecting software development as expressive activity.
The most concrete piece of Byrne’s argument is not doctrinal, it is a case study from his own practice. He describes representing SaSu, a small American mental health forum with no UK presence, which voluntarily geoblocked British users under the UK’s Online Safety Act. Ofcom initially accepted the geoblock as compliance, then reopened the case under political pressure, used VPNs to access the blocked site, cited that access as proof the block failed, fined the site 950,000 pounds, and closed the file 480 days later having collected nothing. Byrne’s point is not that AI poses no risk. It is that regulators, once created, answer to political pressure rather than to the evidence in front of them, and a body built to police frontier AI labs would inherit the same incentives.
Byrne does not dismiss the underlying risk debate; he explicitly rejects the extreme position too, arguing that if AI genuinely posed existential-level danger, the logical response would be to ban computing outright rather than hand oversight to regulators, a position he treats as absurd on its face. His argument is narrower: state control is not a neutral fallback when the stated harm is uncertain, and the institutions being proposed to manage that uncertainty already have a track record.
For an industry now watching Washington, Brussels, and London draft the first generation of AI-specific rules, Byrne’s brief is a preview of the legal fight labs should expect if “safety standard” language gets written into binding law rather than left as a voluntary pledge: expect First Amendment litigation in the US the moment any pacing rule touches what software can be published, not just how it is trained.
Preston Byrne, writing on his personal blog prestonbyrne.com on September 13, 2026.
