AI Rivals Converge on Safety Rhetoric, but Coordination Remains Elusive

A rare show of consensus has emerged among the leaders of the world's most prominent artificial intelligence companies, with Anthropic's Dario Amodei, OpenAI's Sam Altman and SpaceXAI's Elon Musk among those publicly backing a slower, more cautious pace for AI development. The shift follows the resignation of an Anthropic researcher who warned starkly about the risks the technology poses to humanity.

Yet translating that shared concern into practice faces formidable barriers. Domestic and international competition, commercial incentives and resistance from the Trump administration all complicate any coordinated effort to pace AI's advance. The White House has consistently favoured a light-touch regulatory approach, hoping innovation will flourish and keep the United States ahead of China.

In theory, leading AI developers could throttle their models at any time. In practice, breakthroughs have been driven by rivalry to build ever more powerful systems, and both Anthropic and OpenAI are racing toward potentially record-breaking public offerings. OpenAI has said it is holding off on going public this year while it continues safety work.

Amodei recently published a lengthy essay outlining a plan to slow advanced AI development. One proposal would have frontier labs grant independent outside evaluators "ongoing, employee-like access," complete with offices, badges and company laptops. Anthropic is unilaterally committing to the step, and Altman called it a "great idea" on social media, saying OpenAI will follow.

Altman also welcomed a federal framework for safety standards but said the company need not wait for legislation or an antitrust exemption. Amodei's essay calls for government regulation and coordination among frontier labs, supported by the U.S. and other democratic governments, to establish common safety standards and limits on unchecked progress. He acknowledged the difficulty of coordinating with authoritarian governments, specifically citing China.

His plan sketches tiers of international agreement. The most feasible would ban narrow but obviously dangerous uses, such as producing biological weapons. Harder would be pre-release testing through a possible global standards body for cybersecurity or biological risks. Hardest would be a "speed limit" on models capable of recursive self-improvement, or a full pause. Amodei compared capping missile numbers in Cold War treaties to limiting AI's destructive potential while preserving deterrence.

Altman wrote that "pacing" does not mean stopping, and that progress should be slower than it otherwise could be, with safety cases and monitoring carrying significant costs. "No amount of American competitive pressure should justify recklessness," he said.

Experts remain divided on feasibility. Sandra Wachter of the Oxford Internet Institute called a slowdown possible in theory but "highly unrealistic" given the state of the world, suggesting governments could hold companies accountable and restrict access to electricity, water and other data-centre resources. Aidan Gomez, co-founder of the Canadian lab Cohere, warned that embedding evaluators could concentrate power among a handful of labs in one country and objected to antitrust waivers for frontier companies. Elham Tabassi of the Brookings Institution said independent auditors remain company-controlled until commitments are solidified and public, stressing the need for scientifically valid testing.

Zahra Timsah of governance platform i-GENTIC AI noted that within days competing CEOs found common ground and senators from both parties advanced oversight proposals. "It's receiving attention, but attention is not the same as implementation," she said.