Four of the most powerful figures in artificial intelligence agreed over the weekend to slow down. Sam Altman of OpenAI, Dario Amodei of Anthropic, Demis Hassabis of Google DeepMind, and Elon Musk of SpaceX endorsed a proposal to "pace the frontier," calling for third-party auditors embedded in labs, domestic regulation, and a global agreement to throttle development speed. The pledge lasted about as long as a news cycle before critics labeled it a cartel in disguise.

The proposal and its immediate backlash

Amodei laid out the case in a recent essay, arguing that AI has been advancing drastically faster since roughly this summer. His three-step proposal asks labs to allow external auditors like METR, Apollo Research, and Redwood Research to embed within their organizations, flag problematic findings, and potentially blow the whistle on unsafe practices. He also called for speed limits on recursive self-improvement, comparing the concept to caps on missile stockpiles.

Skeptics pounced immediately. The complaint is not new: Big Tech companies have spent years getting ahead of regulation by lobbying for their own preferred rules or promising self-regulation that never materializes. Social media platforms ran the same playbook a decade ago, proposing policies that hit smaller competitors harder while wrapping self-serving goals in altruistic language. The term "safety-washing" has become shorthand for making meaningless changes that give the false impression of actual safeguards.

Industry insiders are split on sincerity

People across the AI ecosystem told The Verge they believe the verbal agreement is a step in the right direction, but not close to sufficient. Nick Reese, an adjunct professor at NYU and the former director of emerging tech policy at the Department of Homeland Security, said he does not think it is all hot air. Apollo Research CEO Marius Hobbhahn called it one of the best things for safety in a long time if it actually happens. Redwood Research CEO Buck Shlegeris said he is feeling cautiously optimistic.

Everyone qualified their praise with the same caveat: there is a lot of work between a verbal commitment and an ironclad agreement. Daniel Kokotajlo, who leads the AI Futures Project after leaving OpenAI, pointed out that people outside the companies have been calling for this for years. A public letter signed by more than 1,000 AI lab employees in July called for a slowdown after an incident involving OpenAI and Hugging Face. Now the CEOs are bowing to that pressure and claiming credit for it, Kokotajlo said, though not rightfully.

The safety researchers who warned about this first

The recent panic crystallized around Jacob Coxon, an Anthropic researcher who resigned and posted a public letter explaining his choice. The letter has been viewed more than 170 million times on X. Coxon wrote that the people building AI earnestly believe it could kill us all by the end of the decade, and that neither OpenAI nor Anthropic is acting responsibly, instead racing straight to self-improving superintelligence and gambling with our lives.

Other researchers across frontier labs echoed his alarm. Jasmine Wang, an OpenAI researcher, wrote that it is hard to overstate how dangerous it is to speed towards recursive self-improvement. Evan Hubinger, an Anthropic team lead, said on X that Coxon is correct and that researchers really do believe AI could kill all humans, placing the probability at greater than 10 percent over the next decade. Alex Turner, a former Google DeepMind employee, said many researchers believe they are building something that could kill everyone on the planet and that it was literally his day job to think about how to stop that.

Micah Carroll, an OpenAI researcher, described Coxon's position as cross-partisan across research teams at all frontier AI companies. He added that business-as-usual development poses unacceptable catastrophic risk, but warned against hypostitioning catastrophic risks into existence. Carroll argued that safety requirements with teeth, international coordination, and a consensus to not build artificial superintelligence without sufficient safety advances could greatly reduce those risks.

Recursive self-improvement is the milestone driving the fear

The technical concept at the center of everything is recursive self-improvement, or RSI. This refers to a point at which AI models can train, advance, and create new versions of themselves without human involvement. Anthropic has said this could arrive as early as early 2027. OpenAI's chief scientist wrote earlier this month that the company is directing significant resources toward reaching this goal.

Researchers increasingly worry that RSI precedes a cascade of more serious concerns, including cybersecurity incidents that impact society at large. Amodei cited RSI as a major factor in his decision to call for a slowdown, writing that left unchecked it could outrun our ability to understand and control these systems and must be pursued very carefully, if at all.

China complicates everything

The single biggest obstacle to an AI slowdown has a name: China. AI leaders and politicians have long positioned China as the reason US progress cannot slow down, arguing that no matter how dangerous the technology becomes, they would rather it be in American hands. Chinese Foreign Ministry spokesperson Guo Jiakun pushed back on the calls for a slowdown on Monday, calling it fearmongering.

The China argument has become a reflexive defense. Sacha Haworth, executive director of the Tech Oversight Project, said China gets brought up as a bogeyman every time an industry wants to escape oversight. The Midas Project's Tyler Johnston and Redwood Research's Shlegeris both said that even without Chinese cooperation, US coordination is still vital. Johnston compared the situation to nuclear de-proliferation, noting that it was in both the US and Russia's interest to coordinate on that issue.

Regulation under Trump is a non-starter

President Trump posted on Monday that the only guardrails AI needs is a strong and smart president, and that the US has that in spades. He also called Nvidia CEO Jensen Huang while Huang was onstage at a conference, telling the crowd on speakerphone that recent concerns about AI were a hoax and that the robots will not be taking over. The administration's posture leaves voluntary agreements as the only viable near-term option, which is precisely why critics worry about regulatory capture.

Reese compared the situation to a cartoon of people in a car driving off a cliff while celebrating that they are ahead of China. He said the industry has never been in a situation where it actually understood what it was driving towards regarding AI, always racing toward an amorphous undefined end state without understanding the path or the finish line. That, more than any proposed slowdown, may be the real problem.