The companies’ endorsement shifts the AI safety debate from broad concern to a harder policy problem: who decides when progress is too fast, and what can governments actually do about it?
OpenAI and Anthropic have given Washington a problem that is easier to state than to administer: when, if ever, should government try to pace the AI race? The rival labs endorsed a letter signed by hundreds of AI company employees calling on U.S. officials to prepare ways to temper AI progress if it accelerates too quickly, including through diplomacy with other countries, The Washington Post reported July 29 from San Francisco.
The move does not answer how a limit would be measured, enforced or reconciled with the companies’ own competition to build stronger systems. It does, however, move the argument from whether AI safety should be taken seriously to what kind of public authority could act before the market sets the pace on its own.
The central question is speed
The endorsement is not described as a call to halt AI development now. Its narrower point is conditional: if progress becomes too rapid, governments should have a way to slow it.
That distinction matters. A standing pause would be one kind of policy. A pacing system would require triggers, evidence and a decision-maker able to say when development or deployment has crossed a line.
OpenAI and Anthropic are not outside observers of that system. They are major participants in the frontier AI race, which makes their support notable and also complicated. Any rule strong enough to affect the industry would likely affect them too.
Employees are pressing from inside the industry
The letter’s base of support came from hundreds of workers at AI companies, according to The Washington Post. Their appeal reflects concern from people close to the technology rather than only from academics, activists or regulators.
The concern is partly about incentives. Companies compete for customers, talent, infrastructure and investor confidence. Faster releases can produce an advantage, while slower review can create time for testing, outside scrutiny and public debate.
That is the tension behind the word pacing. It suggests that the release rhythm for powerful systems should not be set only by competitive pressure. The unresolved issue is what proof would justify slowing that rhythm and who would be trusted to make the call.
Washington already has a safety channel
The federal government is not starting from zero. In 2024, the National Institute of Standards and Technology announced agreements with OpenAI and Anthropic for AI safety research, testing and evaluation through the U.S. AI Safety Institute.
NIST said the agreements created a framework for the institute to receive access to major new models from each company before and after public release. The work was aimed at evaluating capabilities, studying safety risks and exploring ways to reduce those risks.
Elizabeth Kelly, then director of the U.S. AI Safety Institute, said safety was essential to technological innovation and called the agreements an important milestone. The institute also planned to coordinate with the U.K. AI Safety Institute.
Those arrangements were framed as technical collaboration. The new endorsement raises a separate policy question: whether voluntary safety cooperation could become something closer to government influence over the pace of AI deployment.
The approval line remains sensitive
One reason the debate is difficult is that slowing AI can sound close to requiring permission before new systems are released. That is a line some industry leaders have treated carefully.
Reuters reported in June that OpenAI CEO Sam Altman was expected to urge U.S. lawmakers not to require AI developers to obtain government approval before releasing new models. That position does not rule out all oversight, but it shows the difference between safety review and pre-release licensing.
There is also a competition concern. Large companies can often absorb compliance systems, reporting duties, evaluations and security requirements more easily than smaller rivals. Critics may therefore ask whether new rules would protect the public, strengthen incumbents, or do both.
That skepticism does not negate the safety argument. It does mean any pacing mechanism would have to be judged not only by its stated goal, but also by who gains power from the way it is designed.
A U.S.-only answer may be too weak
The letter’s call for international diplomacy points to a practical problem: AI development does not stop at national borders. Labs, researchers, cloud infrastructure, chips and customers are part of a global system.
If the United States slowed only domestic companies, critics could argue that foreign competitors would benefit. If the United States sought broader limits, governments would need shared definitions, verification methods and enforcement tools.
That is far harder than announcing a principle. AI is software, compute, data, research practice and commercial deployment at the same time. A workable framework would have to account for private labs, open-source models, academic work and companies operating across jurisdictions.
The policy design is still missing
The endorsement gives policymakers a clearer mandate to examine AI pacing, but it does not supply the machinery. Possible approaches could include mandatory safety evaluations, compute thresholds, incident reporting, coordinated release delays, international standards or emergency powers tied to specific model capabilities.
Each option carries trade-offs. A light system may not change company behavior. A heavy system may slow useful products, concentrate power among large firms or give agencies authority they are not ready to use.
The immediate takeaway is not that an AI speed limit now exists. It is that two leading AI companies have backed a call for government to consider when one might be needed, leaving Washington with the hardest questions still unanswered.
Leave a Reply