Anthropic CEO Dario Amodei has called for frontier AI companies to ensure that advances in model capabilities do not move faster than the safety work needed to understand and manage them.
In his new essay, “We Must Pace the Frontier,” Amodei argues that companies developing the world’s most advanced AI models need stronger mechanisms for independent oversight as these systems become increasingly capable.
The argument adds another question to the AI race.
It is no longer only about who can build the most powerful model, but also whether safety practices can keep pace as those models become more capable.
What does “frontier AI” mean?
Frontier AI generally refers to the most advanced and capable AI systems currently being developed, including the leading models produced by companies such as Anthropic, OpenAI and Google DeepMind.
Amodei is not calling for AI development to stop.
Instead, he argues that the pace of capability development should remain compatible with the industry’s ability to evaluate risks and improve safety.
Amodei proposes three stages
His proposal includes:
- Permanent, independent third-party evaluators within frontier AI companies
- Common safety standards among AI companies in democratic countries
- Longer-term global coordination around advanced AI development
Anthropic is also taking a step towards the first proposal.
The company says independent evaluators will receive continuous, employee-like access, allowing them to examine safety practices, incidents and work designed to ensure AI systems behave safely and remain aligned with human intentions.
OpenAI responds
The discussion quickly extended beyond Anthropic.
OpenAI CEO Sam Altman responded to Amodei’s announcement by supporting the approach and saying OpenAI would also provide independent evaluators with similar access.
The response is notable because it suggests that independent verification could increasingly become part of how frontier AI companies demonstrate the safety of their systems.
From model capability to verifiable safety
Speed and model capability remain central to competition between frontier AI companies.
But another measure may be emerging alongside them.
Not only how powerful an AI system is, but how independently and transparently its safety can be verified.
If independent evaluation becomes a broader industry practice, competition at the frontier of AI may increasingly involve not only model performance, but also the credibility of the safety systems surrounding it.
For governments, businesses and organisations adopting increasingly capable AI systems, that distinction could become increasingly important.
Source
Dario Amodei, Anthropic CEO, “We Must Pace the Frontier”
Read the original essay
