Dario Amodei proposed three mechanisms to “slow the pace of progress” in artificial intelligence models, including independent monitors within companies, shared safety standards, and international coordination. Anthropic, for its part, is committing to providing expanded access to external evaluators, while the feasibility of the proposals and their potential conflict with competition laws remain open questions.
Dario Amodei, CEO of Anthropic, called for slowing the pace of improvements in the capabilities of artificial intelligence models, proposing three approaches to address the growing risks associated with what industry companies call the technology’s “advanced frontier.” Amodei said Anthropic would unilaterally commit to one of these approaches, calling on governments to require other companies to follow it.
The proposal came in a new post amid intensifying debate over AI safety, and after warnings from researchers within the industry about the speed of development and the growing systems’ ability to help build the next generation of AI models. Amodei said that the OpenAI-HuggingFace breach, along with the major acceleration in model capabilities in recent months, had convinced him of the need for a more cautious approach.
Independent Monitors Within AI Companies
The first proposal is to bring external evaluators, such as METR, into advanced AI companies. Amodei likens this arrangement to having monitors or regulatory bodies within financial institutions, allowing evaluators to verify companies’ compliance with their commitments regarding slowing development and safety, and to ensure that safety incidents are reported.
Anthropic says it will unilaterally commit to this arrangement. Evaluators would receive access badges, offices, and computers, along with access to information close in scope to that available to internal risk-assessment teams, except where prohibited by law or contracts. The proposal follows criticism directed at OpenAI for failing to report an incident in which AI agents took over a model on a German wiki website, according to the article.
Shared Standards and Limits on the Speed of Development
The second approach involves coordination among leading companies based in democratic countries to establish shared safety standards, along with limits on the rate of unsupervised progress. Amodei believes this coordination could face practical difficulties, particularly amid competition between companies and concerns that any coordinated halt or slowdown could be interpreted as a violation of antitrust laws.
To address this concern, he proposed that the U.S. government mediate or enable these discussions through a limited legal exemption allowing certain safety-related discussions. The proposal, as described in the article, does not call for full government involvement in managing companies, but for providing a legal framework for safety discussions among competitors.
China and International Coordination
Amodei acknowledged that the call to slow development faces the common argument that any U.S. slowdown could give China an advantage in AI. But he said the United States and technology companies could slow China’s progress enough to widen the gap over the next three to five years by preventing the sale of powerful chips and semiconductor manufacturing equipment to Chinese companies, and by tightening restrictions on model-distillation technologies.
The third approach is global coordination involving the United States and its allies, with an attempt to cooperate with authoritarian governments, including China, within narrow limits. Amodei acknowledges that the chances of reaching an agreement are limited, but sees room for agreements banning specific and clearly dangerous uses, such as employing AI to produce biological weapons or enable users to do so.
What Really Matters?
What stands out in the proposal is not merely a general call for “AI safety,” but an attempt to turn it into verifiable mechanisms: independent monitors, reporting commitments, shared standards, and limits on the speed of development. If the proposal concerning evaluators is implemented, it would give external entities access to risk-assessment processes inside a leading company, potentially raising the level of scrutiny compared with voluntary commitments alone.
At the same time, the post offers no final solution to the conflict-of-interest problem. Companies competing to develop models would need to exchange safety information without revealing what they regard as trade secrets, while coordination among them could lead to legal scrutiny under competition rules. Disagreement over the nature of the risks also remains: some critics question long-term catastrophic warnings and believe they may divert attention from current harms, while others warn that regulatory proposals could ultimately serve major companies and entrench their influence.
Amodei emphasizes that he does not reject the benefits of AI, but believes that improving quality of life will be achieved only by building the technology properly and carefully using the time provided by a slowdown. However, the source does not explain how the “acceptable pace” of progress will be measured, who will determine which incidents must be reported, or how the standards will be applied to companies that do not join this framework.