Leading AI laboratories have called for independent oversight mechanisms and government-imposed rules to slow the pace of artificial intelligence development. However, critics warn that such measures could inadvertently create a regulatory cartel while raising barriers to entry for startups.
Dario Amodei, chief executive of Anthropic, published “We Must Pace the Frontier,” urging stricter controls on AI advancement. Sam Altman of OpenAI publicly endorsed the proposal but did not announce any immediate pause in development at his company. Both executives emphasized the need for independent inspectors, industry standards, and government coordination.
Amodei contends that current AI systems are improving at a rate that outpaces human engineering capabilities. He cites incidents where swarms of AI agents conducted unauthorized cyberattacks, warning that such threats could escalate to threaten the internet on a scale measured in hundreds of billions of dollars within months. His proposed solutions include permanent evaluators with employee-level access within major labs, standardized protocols among democratic nations (brokered by Washington under an antitrust waiver), and mechanisms for cooperation with China.
Altman agreed to the concept of pacing the frontier and stated that OpenAI would allow independent evaluators. However, Amodei remains the primary architect of the broader framework involving standards, checkpoints, and official oversight. When two dominant firms jointly request state involvement in rule-making, critics question whose interests those rules will ultimately serve.
David Sacks, co-chair of President Trump’s Council of Advisors on Science and Technology, directly addressed the proposal. Sacks argued that Anthropic and OpenAI already function as a duopoly by market share, revenue growth, and model capability. “Go ahead,” he wrote. “You guys are the frontier.” He noted that the simplest way to prevent superintelligence is for these companies to refrain from building it altogether. Demanding a specific regulatory framework as a condition for restraint, he said, would appear as blackmail, while refusing to slow down without such an agreement would indicate another instance of regulatory capture.
If Amodei’s risks are real, Anthropic could reduce its own contributions without waiting for competitors or Beijing. Permanent inspectors represent a cost burden that startups cannot easily absorb but incumbents can leverage. Release checkpoints function similarly to licenses that favor companies already equipped to navigate them. A pace negotiated under an antitrust waiver is likened to a cartel operating under the guise of safety protocols.
China poses a critical verification challenge: if democracies slow AI development and Beijing defects, it could hold the most powerful technology in history. Unlike physical weapons, advanced AI models can be stolen, copied, and imitated without clear verification methods. Amodei described “pacing the frontier” as an honor system that would bind only those deemed honorable while leaving others behind.
Allum Bokhari, director of the Foundation for Freedom Online, argued that if artificial intelligence is genuinely dangerous, the United States must lead in regulation rather than cede the field to authoritarian adversaries. He noted that veterans of past “trust and safety” offices have transitioned into AI safety roles, where “safety” often translates to machine learning fairness—constraining models to avoid discriminatory outcomes. When Amodei seeks inspectors with authority to publish findings within major labs, Bokhari questioned whether those inspectors would audit for rogue systems or decide which claims are permitted online.
Sacks has previously accused Anthropic of regulatory capture through fear tactics and advancing “woke AI” via blue-state legislation. This summer, an industry letter endorsed open-weight models that anyone can download and run; Meta signed on but Anthropic did not. Amodei has also called for a crackdown on industrial-scale distillation—a practice where smaller labs close the gap by querying larger systems.
Anthropic’s unilateral inspector program remains largely under its control. The steps required to bind it would necessitate collaboration with Washington and Beijing, a process that takes time. Meanwhile, Bernie Sanders has announced a permanent ban on “superintelligence” and a pause on advanced development. The slowdown coalition spans from democratic socialists to the industry’s wealthiest executives. They share a common goal: to slow AI development and expand state oversight. However, their disagreements center on who will sit on the regulatory commission.
Critics note that if advanced AI is dangerous, liability can be clearly defined. Yet, when it comes to power, dispersing its development among multiple builders presents a different model than concentrating it within two laboratories under government-approved inspectors. The laboratories argue the machines must slow down. And in doing so, the process they are designing would also slow down anyone building outside their control.
