AI leaders back slowdown after safety warning from Anthropic researcher
Top leaders in the artificial intelligence industry, who often disagree on major issues, have in recent days backed calls to slow the development of AI, putting safety at the centre of the debate. The rare alignment among Anthropic's Dario Amodei, OpenAI's Sam Altman, SpaceXAI's Elon Musk and others came after an Anthropic researcher resigned and warned of the threat the technology could pose to humanity.
Even so, a coordinated slowdown faces major obstacles, including competition within the US and globally, the drive for profits, and resistance from President Donald Trump's administration. Experts say the moment could still push the industry to rethink what progress should look like. Nick Reese, a former director of emerging technology policy at the Department of Homeland Security and an adjunct professor at New York University, said: "What if we considered success in AI the same way that we consider success for airlines? On time and haven't had a crash in who knows how long. We don't tolerate commercial airline crashes, but we do tolerate significant AI failures."
In theory, the leading AI companies could slow work on their models at any time. In practice, however, recent advances have been driven in part by fierce competition to build more powerful systems, with support from a White House keen that the US should not lose ground to China. Anthropic and OpenAI are also moving towards potentially major public offerings, though OpenAI has said it is not going public this year as it continues its safety efforts.
For that reason, the proposals now being discussed depend on cooperation between companies and, to some extent, regulation by governments. In a recent essay, Amodei set out a plan to slow advanced AI development. One proposal calls for leading AI firms, often described as frontier labs, to give independent outside evaluators "ongoing, employee-like access" so they can monitor safety practices. The evaluators, under the proposal, would have offices, access badges and company laptops. Amodei said Anthropic was "unilaterally committing" to that step, and Altman called it a "great idea" on social media, saying OpenAI would do the same.
Altman said he would welcome a "federal framework" setting safety standards for advanced AI, but added that his company did not need to wait for legislation or an antitrust exemption before acting on AI safety. Amodei also called for coordination among frontier AI companies with support from the US government and other democratic countries. He said he hoped the leading labs could create "common safety standards as well as limits on the rate of unchecked AI progress".
Amodei's plan also includes efforts by the US and other countries to coordinate with "authoritarian governments", though he acknowledged the difficulty of that, especially in securing cooperation from China. He outlined several levels of possible global agreement. The most feasible, he said, would be a ban on "certain narrow and obviously dangerous uses of AI" such as producing biological weapons or helping users do so. A more difficult step would be persuading the US and its adversaries to test models before release, possibly through a global standards body, for risks linked to cybersecurity or biological threats. Harder still would be a "speed limit" on systems that can improve themselves, a process known as recursive self-improvement.
Arguing for a slower pace, Amodei wrote: "Slowing the rate from extremely fast' to only somewhat fast' gives up relatively little strategic advantage, while potentially greatly improving safety." He compared it to Cold War treaties in which capping missiles limited destruction while preserving deterrence. He said the hardest proposal to put in place would be a "full pacing, or even pause,'" under which participating governments would substantially limit the overall rate of AI development. Altman, writing on X, said pacing did not mean stopping. "But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs," he said. "Pacing will be well worth this cost," he added, saying "no amount of American competitive pressure should justify recklessness."
Experts said the growing agreement within the industry was notable, but they also pointed to unresolved questions over whether any slowdown could actually happen. Sandra Wachter, a professor of technology and regulation at the Oxford Internet Institute, said a slowdown was possible "in theory". But in practice, she said, it would require major coordination between companies and governments across countries, which she called "highly unrealistic" given "the current state of the world". She said governments could try to promote a slowdown by holding AI firms more accountable for risks and by regulating their access to electricity, water and other resources used to run data centres.
The idea of placing outside evaluators inside AI companies has largely been welcomed in the industry, but some critics have questioned how independent such assessors could really be and how standards would be set. Cohere co-founder and chief executive Aidan Gomez said the result could be greater concentration of power. "A handful of the most powerful labs based in one country would agree on shared standards and the limits to how fast the technology should advance," he said. He also rejected Amodei's suggestion that the US government could give frontier labs waivers from antitrust restrictions. "If this is the most consequential technology in human history, then the rules for it cannot be written by a small group of commercially aligned companies behind an antitrust waiver," Gomez said.
Elham Tabassi, director of the AI and Emerging Technology Initiative and senior fellow at the Brookings Institution, said that until the measures companies are promising are formalised and made public, independent auditors would remain a "voluntarily-provided, company-controlled" step. She said AI companies must continue investing in "scientifically valid" methods to test and measure their models as they pursue greater safety, adding that regulation would be weaker without such tools.
The Trump administration has long backed a light-touch approach to AI in the belief that innovation would flourish and that the US would move ahead of China. Experts said the rivalry between American and Chinese AI companies has increasingly been compared to an arms race, making collaboration more difficult. Reese said this creates "conditions where now we have this existential adversary". Even so, some industry figures remain "cautiously hopeful" that AI development could move in a more safety-focused direction. Zahra Timsah, co-founder and chief executive of governance platform i-GENTIC AI, said: "Within days, you have CEOs that are competing - all of them, all of a sudden, found common ground. You have senators from both parties that started to advance different forms of oversight. It's receiving attention, but attention is not the same as implementation. Will it really work? I don't know."
For now, the debate has brought unusual agreement among AI leaders and fresh attention to safety, but questions remain over how any common standards, oversight or slowdown would be enforced in the face of political resistance, commercial pressure and global competition.
