The AI Slowdown Pact Is a Power Play Disguised as Safety

The Pact That Looks Like a Fence

In a weekend agreement that drew immediate suspicion, Sam Altman, Dario Amodei, Demis Hassabis, and Elon Musk pledged to “pace the frontier” of AI development. The proposal, fleshed out in Amodei’s essay, called for third-party auditors inside labs, domestic regulation, and a global slowdown deal. On paper, it reads like a responsible industry maturing. In practice, it reads like a moat.

Skeptics didn’t wait long to name it: a cartel. The four CEOs control roughly the most compute-intensive frontier labs on Earth. If they collectively decide to slow down — while embedding external auditors of their choosing and pushing for rules that apply mainly to labs of their scale — they don’t just manage risk. They manage competition.

The open-source movement felt the sting immediately. Smaller labs and independent developers have spent years building models that pressure the frontier. A voluntary slowdown targeting only the biggest players sounds equitable until you notice who defined the threshold and who picks the auditors.

Safety Is Real. The Incentives Are Messier.

The alarm inside AI labs is not manufactured. Researchers at Anthropic, OpenAI, and Google DeepMind have warned, sometimes publicly, that recursive self-improvement — the point where models can train and advance themselves without human involvement — may be closer than anyone admits. Anthropic has suggested it could arrive as early as early 2027. OpenAI’s chief scientist confirmed the company is directing significant resources toward that milestone. Jacob Coxon’s resignation letter from Anthropic, viewed more than 170 million times on X, described a culture racing toward self-improving superintelligence while treating catastrophic risk as collateral.

That context matters. These are not executives who discovered fear this weekend. They are executives who have been hearing the same warnings from their own researchers for months — warnings that Coxon’s letter made impossible to ignore.

Daniel Kokotajlo, who leads the AI Futures Project and previously worked at OpenAI, put it plainly: people outside these companies have been calling for exactly this kind of slowdown for years. More than 1,000 AI lab employees signed a public letter in July demanding a pause after the OpenAI–Hugging Face incident. The CEOs are now adopting those demands and attaching their names to them.

Whether that is genuine alignment with safety research or strategic co-optation depends on what comes next. Voluntary frameworks have a poor track record. Nick Reese, formerly the Department of Homeland Security’s director of emerging tech policy and now an adjunct professor at NYU, compared the industry’s move to the social media playbook from a decade ago: propose your own rules before someone else writes worse ones for you.

The Regulatory Vacuum Makes Everything Easier

The timing is hardly accidental. Under the current administration, comprehensive AI regulation is unlikely. Trump has publicly dismissed AI risk as a “hoax,” called recent safety concerns fabricated, and told Nvidia’s Jensen Huang on speakerphone at a conference that the only guardrail AI needs is a strong president. The White House has agreed to limited pre-release review periods for models, but that is a narrow concession, not a regime.

In that environment, industry self-governance is the default. That makes the cartel question unavoidable. When the fox proposes the henhouse rules, regulatory capture isn’t a conspiracy theory — it’s the baseline assumption.

Sacha Haworth, executive director of the Tech Oversight Project, did not mince words. She called voluntary frameworks essentially regulatory capture and argued that Congress should not outsource responsibility to the industry again. Daniel Lobo-Lewis of the Political Integrity Project predicted the same fate as Meta’s Oversight Board: a body with enough prestige to seem credible and enough powerlessness to change nothing.

The China Card

No discussion of AI governance omits China. It is the oldest argument in the book and the most durable. Politicians and executives alike invoke Beijing as proof that the U.S. cannot afford to slow down. One X post captured the sentiment literally: “If I am going to die at the hands of killer AI, I want it to be American, not Chinese.”

Chinese officials pushed back. Foreign Ministry spokesperson Guo Jiakun called the slowdown rhetoric “fearmongering.” The implication was clear: Beijing will not cooperate, so the U.S. should keep racing.

But several experts pushed back on that framing. Tyler Johnston of the Midas Project noted that assuming China won’t coordinate is itself an assumption. Buck Shlegeris of Redwood Research argued that reckless development serves neither government, and that international coordination on existential-risk technology is far from unprecedented — nuclear deproliferation involved adversarial states finding ways to cooperate when the alternative was mutual destruction.

Haworth went further, calling the China argument a familiar escape hatch. “China gets brought up as a bogeyman every time that an industry wants to escape oversight,” she said. Kokotajlo recalled a cartoon of people in a car heading off a cliff, shouting, “Hooray, we’re ahead of China.”

The point is not that China cooperation is guaranteed. It is that theChina card has been used to justify almost every delay in oversight, and it will continue to be used unless something stronger replaces it.

Who Wins, Who Loses

If the pact materializes as written, the winners are the labs that already sit at the frontier with the most compute and the deepest pockets. They gain a veneer of responsibility, an excuse to slow competitors who cannot afford the same audit infrastructure, and a head start in shaping whatever norms eventually harden into law.

The losers are open-source developers, smaller labs, and anyone outside the narrow circle of frontier companies that get to define what “responsible” means. They are also the people who benefit from speed — researchers trying to steer alignment work, engineers building safeguards, journalists and policymakers trying to keep up.

There is a middle path that some experts have proposed, and it is worth watching: requiring labs to disclose and reduce their compute budgets for frontier research, with third-party auditors verifying the numbers. Compute is the closest thing the industry has to a throttle. If you can slow compute, you slow progress in a measurable way. If you can measure it, you can enforce it.

That is still a proposal, not a policy. The current agreement is verbal and voluntary. It is a good sign if it pushes the conversation toward enforceable limits. It is a bad sign if it becomes the final word.

What Comes Next

The most important question is not whether the pact is sincere. Sincerity is hard to prove and easier to fake. The question is whether the industry can be held to something stronger than a weekend statement.

Reese captured the structural problem: no one in power has a realistic vision for where this is going. “We’ve always had this amorphous undefined end state,” he said. “It’s really hard to race when we don’t even understand the path or even understand what the finish line is — or even if there is one.”

That ambiguity is exactly where cartels thrive. Without a clear public standard, the people with the most compute set the pace. With one, they can be held accountable when they miss it.

Regulators in the U.S. and abroad should treat the current proposal as a starting point, not a settlement. The framework Amodei outlined contains useful elements — independent auditors, whistleblower protections, measurable constraints. But history shows that voluntary commitments from industries with market-power incentives tend to soften under pressure. The next administration may not share Trump’s dismissal of AI risk. If a Democratic president takes office and faces the same escalation in internal warnings, the choice will be between a weak voluntary regime and actual law. The industry’s current move is best understood as a bid to make the weak version permanent before anyone asks for the strong one.

The slowdown could be genuine. It could also be the most expensive fence ever built around a race — one that keeps everyone else out while the incumbents decide how fast to run.