AI leaders call for a slowdown. Trump pushes back.

The central AI story this week is a split over whether the frontier should slow down. Anthropic's Dario Amodei called for coordinated restraint on September 12; OpenAI's Sam Altman and Elon Musk backed the message. President Donald Trump pushed back on September 13, emphasizing competition with China. That confrontation belongs at the center of our coverage, including today's developments in model safety. Bloomberg's reporting establishes the timeline.

In We Must Pace the Frontier, Amodei argues that safety work needs more time to catch up with increasing capabilities. His proposal has three parts: embedded outside evaluators, coordination among companies in democratic countries, and international cooperation. Anthropic commits to the evaluators first. He describes slowing capability advances while continuing technical work, rather than an immediate halt to all training. This is a proposal from an industry participant with commercial interests, not an agreed international safety regime.

Trump's response put maintaining the US lead first and dismissed some of the warnings. The administration's position is more specific than a ban on companies slowing themselves: WIRED reports that adviser David Sacks said labs could act voluntarily while rejecting their case for government-enabled coordination. The distinction matters: individual restraint, an industry agreement, and an enforceable public rule are different commitments.

The September 16 development is OpenAI's new misalignment disclosure framework, accompanied by six reports from training or evaluation. OpenAI says these examples do not establish how often misalignment occurs and are not a complete inventory. They supply concrete material for the slowdown debate, but publication alone does not demonstrate that the underlying risks are controlled.

Our view: warnings from the companies building these systems deserve scrutiny and action, especially when the public bears risks it did not choose. So do the proposed remedies. We should ask who selects evaluators, whether findings can be published freely, what would actually trigger a delay, and who can enforce one. Support for oversight should not depend on taking executives' assurances at face value. Enthusiasm for useful AI is compatible with demanding that its developers stop when safety cannot keep up.

Updated September 17, 2026: The earlier version led with OpenAI's disclosure framework and omitted the broader slowdown dispute. This edition has been revised to address that omission. The CEOs' calls and Trump's response occurred September 12–13, not today.