More than 1,100 employees of OpenAI, Anthropic, Google, Meta and other technology firms have signed a declaration asking the Trump administration to help slow the pace of automated AI research, warning that systems able to improve themselves could soon outrun human oversight. The statement asks the United States to join a global effort to build governance and technical tools so the frontier of self-improving AI advances "at a deliberate pace." Signatories include Anthropic chief executive Dario Amodei and OpenAI co-founders John Schulman and Wojciech Zaremba, alongside senior staff from Microsoft, Google, Mistral and Thinking Machines.
The core worry is recursive self-improvement: AI systems that help design the next, more capable generation of AI. The signers argue that competitive pressure leaves little time to test or control what results, and that no single company or country can slow down unilaterally without ceding ground. They are not calling for a halt. They want capability gains paced just enough for safety research to keep up.
The declaration formalises an argument Amodei made in a September 12 essay urging the industry to "pace" frontier development. His plan had three parts: independent evaluators given permanent access inside the labs, shared safety standards set with governments, and eventual global coordination. OpenAI's Sam Altman said his company would adopt embedded evaluators too, and Elon Musk's response was blunt: "Dario is right."
That apparent unity has limits. A rival camp, including Meta's Mark Zuckerberg, Nvidia's Jensen Huang and investor David Sacks, argues that slowing down hands the global lead to China and that heavy new rules mainly protect the largest incumbents while crushing smaller and open-source developers who cannot absorb the compliance load. Both sides claim to want safety; they disagree on who should define it, and at what cost.
The labs have already tested a lighter form of cooperation. OpenAI and Anthropic ran safety and misalignment evaluations on each other's models and published separate reports examining alignment, sycophancy and hallucinations. A formal agreement to cross-test reached the contracting stage before it was abandoned, according to reporting summarised on The AI Daily Brief. The commercial reason is plain: giving a competitor privileged access to probe your model exposes its guardrails and failure modes, and query access is the raw material for distilling a cheaper copy. That asymmetry is why voluntary cross-testing may be structurally unstable, and why some backers want a rule imposed on everyone instead.
Not everyone reads the alarm at face value. The Associated Press reported that the warnings coincide with anticipated stock listings on a jittery market and the US midterms, and that the rhetoric lets the labs shape the terms of their own regulation. A Gartner analyst called the joint safety reports "a sophisticated attempt to frame the safety debate on the industry's own terms." Sarah Shoker, who once led OpenAI's geopolitics team, said focusing on existential risk deprioritises present harms: "If you look at the use of AI in military tech, you can see that these systems are already used to kill people." President Donald Trump has dismissed talk of AI risk to humanity as a "HOAX" meant to help China, which leaves the declaration's central request, federal action to slow the race, without an obvious path through the current White House.