Story
September 14, 2026
AI Chiefs Call for a Slowdown, but Critics See the Leaders Protecting Their Lead
Anthropic and allied AI leaders argue that slowing the race is the only way to keep increasingly capable systems under human control; skeptics counter that the companies best placed to set the brakes are also those with the most to gain from them.
The latest safety flare-up began with former Anthropic researcher Jacob Coxon’s public resignation. He accused Anthropic and OpenAI of racing toward self-improving superintelligence while “gambling with our lives,” turning an internal industry anxiety into a highly visible challenge to the frontier labs.1
On Saturday, Anthropic chief executive Dario Amodei answered with a more concrete prescription: “We must slow the pace at which we improve the capabilities of AI models.” His concern is not a distant hypothetical. Amodei warned that, within six to 12 months, a sufficiently capable rogue-agent swarm could potentially seize control of the internet through a persistent botnet and inflict vast damage.2
His three-part plan starts at home. Anthropic says it will give independent evaluators permanent, employee-level access to its systems and internal processes, with scope to report safety incidents. It then calls for common limits among democratic-country labs, backed by government help to clear antitrust obstacles, and eventually for limited global coordination—including with China—on plainly dangerous uses such as biological weapons.3
The proposal quickly drew support from rivals. Elon Musk put it simply: “Dario is right.”
4 Sam Altman likewise said OpenAI must avoid losing control of the future to AI and insisted that increasingly capable systems must remain in service of people.
5
But agreement among the biggest players has sharpened rather than settled the credibility question. David Sacks noted that Amodei and Altman are, by market share, revenue growth and capability, effectively a “duopoly on frontier intelligence”—and told them to go ahead with pacing.
6 Journalist Brian Merchant’s critique is blunter: rules of this kind could chiefly serve Anthropic and OpenAI, “what regulatory capture looks like in action.”7
That is the fault line beneath the new consensus: a demand for real restraint, and a fear that the firms asking for it may get to define the restraint.