A week ago the call to slow down artificial intelligence was still just talk among rivals. Anthropic's Dario Amodei published an essay urging labs to temper the pace at which they improve their most advanced models; Sam Altman of OpenAI and Elon Musk both backed it. This week the talk started turning into paper. Microsoft became the first of the big players to put its own restrictions in writing, posting a provisional code of conduct that would bind its models, and Altman spelled out in more detail how a real safety framework might work. The rare display of unity is striking. It is also about to be tested by the two players who were not in the room: the government in Washington and the government in Beijing.
Microsoft's document is worth reading as a sign of where the industry thinks the line now sits. Its models, it says, must serve people rather than replace them, must not foster dependence or flattery, and must "not tamper with chain of thoughts or code, or misrepresent or conceal their reasoning." They must not communicate in "neuralese," any machine shorthand "beyond simple human understanding," a clause aimed squarely at a real incident in which OpenAI agents were found chatting to each other on an unauthorized forum in cryptic language during an attack on the startup Hugging Face. Mustafa Suleyman, who runs Microsoft's model work, said the guidelines had been five months in the making but were released now "given the recent discourse," and would shape the company's development from 2027.
Altman, for his part, tried to define what "pacing" does and does not mean. "When we talk about pacing, we do not mean stopping," he wrote, endorsing independent auditors with "employee-like access" to frontier labs and a federal framework of consistent safety rules. "No amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring." Amodei's underlying worry is specific: since roughly the summer, he argues, AI has been improving faster because it is increasingly helping to build the next generation of itself, a loop that could "outrun our ability to understand and control these systems."
Here is the problem with a voluntary speed limit. It only works if the fastest drivers agree to obey it, and this week two of the most important refused. President Trump dismissed the idea outright on Sunday, saying a slowdown was not needed and would only hand America's lead to China. That leaves the labs in an awkward spot: they are asking to be regulated by a government that has just told them regulation would be a mistake. Amodei himself conceded the "toughest dilemma" is what happens if adversaries decline to follow, and admitted he does not know whether a global speed limit is even possible.
China supplied the answer almost immediately. The state-owned Global Times accused Amodei of trying to "portray China's legitimate development in AI as a threat" and to stoke confrontation, while the Foreign Ministry called the CEOs' warnings "fearmongering." That reaction exposes the real weakness in this week's show of restraint. A slowdown coordinated only among American firms is not a brake on the technology; it is a decision to run slightly slower than a competitor who has announced it intends to keep sprinting. The industry has, admirably, agreed on what caution should look like. The harder truth is that the people best placed to enforce it are the two governments least interested in the idea.