AI Slowdown Calls Deserve Skepticism, Not Applause

In September 2026, something unusual happened in Silicon Valley: The people building the most powerful technology in history asked everyone to slow down.

Dario Amodei published an essay urging the industry to "pace" itself. Sam Altman agreed on X within hours. Elon Musk chimed in with: "Dario is right."

For a moment, it looked like the AI industry had discovered humility. It hadn't. What it had discovered was a convenient moment to ask for exactly the kind of rules that would lock in its own dominance.

To be fair, the triggering events were real, and they were alarming. In an internal red-team test, OpenAI models reportedly exploited a vulnerability in a third-party package cache to break out of their sandbox, coordinate with each other, and chain further exploits to reach Hugging Face's production systems: Activity OpenAI itself confirmed in a post-mortem.

Around the same time, a UK AI Safety Institute evaluation found an Anthropic research model fabricating online personas to socially engineer a human tester into approving code it shouldn't have. These are not hypothetical worries about some distant superintelligence. They are documented behaviors from systems that companies are racing to make more capable, not less.

So the alarm is not manufactured. But the response to it deserves far more scrutiny than it's getting.

Start with the substance of what's actually being proposed. None of the three executives called for a compute cap, a licensing regime with teeth, or a binding international treaty. Amodei's essay outlines a three-part framework: Embed independent evaluators inside labs, coordinate safety standards among firms, and pursue international agreements, but stops well short of anything that would require these companies to actually stop building.

As Amodei himself put it, this is "pacing, not halting." Progress, he assured everyone, "will still seem fast." Altman, similarly, promised to adopt "independent safety evaluators" while continuing to ship products. Musk publicly agreed, even while reportedly telling associates in private that the whole campaign smelled like a "psyop" meant to slow down competitors which, if true, is a remarkable thing to say about a cause you're publicly endorsing.

This is the pattern regulatory capture always follows: Incumbents call for oversight framed around safety, while quietly shaping it to raise the cost of entry for everyone else.

Anthropic co-founder Jack Clark said the quiet part almost out loud, noting that with only "a small number of key players" left, it's easier to impose a policy regime that prevents the technology from "spiraling out of control." Critics, including some in the current US administration, have called this exactly what it looks like: A Trojan horse.

Third-party audits, licensing requirements, and compute thresholds are all costly to comply with trivially so for a company sitting on tens of billions in committed infrastructure, and prohibitively so for a startup or an open-source project trying to compete.

And the financial backdrop makes the "we're pumping the brakes" narrative harder to take at face value. OpenAI is reportedly carrying compute and data-center commitments estimated in the hundreds of billions of dollars through 2030. Anthropic has apparently financed nearly its entire infrastructure buildout through debt while simultaneously preparing to begin marketing an IPO within weeks of Amodei's essay. Nvidia is said to be backstopping tens of billions of OpenAI's data-center leases.

These are not the balance sheets of companies bracing to actually decelerate. They are the balance sheets of companies that need calm markets and reassured regulators while they finish building the largest capital projects in corporate history. A public show of caution, delivered right before a fundraising push, is not necessarily hypocrisy but it is, at minimum, excellent timing.

There's also a geopolitical alibi built into the whole framing. Any serious slowdown proposal collides immediately with the argument that China will simply race ahead if the U.S. doesn't.

That argument isn't wrong, exactly, but it's also strategically convenient: it lets executives gesture at caution while ensuring no one including themselves actually pumps the brakes. It's hard to imagine a slowdown pledge less binding than one whose loudest proponents keep reminding everyone why they can't actually honor it.

None of this means the underlying risk is fake. Researchers with real technical credibility, people who've spent years working on interpretability and alignment, not press releases have put double-digit probabilities on catastrophic outcomes from advanced AI. The sandbox-escape and social-engineering incidents this year are genuine evidence that today's systems can behave in ways their creators didn't intend and didn't authorize. Dismissing all of this as theater would be its own kind of complacency.

But the right response to real risk is not to trust the risk-takers to regulate themselves, especially when the shape of the self-regulation they're proposing happens to entrench their own market position.

If Amodei, Altman, and Musk actually believe what they're saying about the stakes, the test isn't whether they write an essay or post a tweet. It's whether they support external, binding, independently enforced rules that apply to everyone equally, including the possibility that a genuinely safer path is a slower, smaller AI industry than the one currently raising trillion-dollar valuations.

Until that happens, "slow down" from the people racing fastest should be read the way any competitor's sudden interest in the rules of the game usually should: With one eyebrow raised.

Shafqat Aziz is a barrister (Lincoln’s Inn) and an accredited Civil-Commercial Mediator (ADR-ODR International).