Dario Amodei proposes external evaluators, common standards, and international coordination to reduce risks associated with the accelerated advancement of artificial intelligence models.
The race to develop increasingly capable artificial intelligence models is beginning to incorporate a variable that had previously occupied a secondary place: the speed of technological progress.
Dario Amodei, CEO of Anthropic, argues that leading AI companies should deliberately slow the pace at which they increase their models’ capabilities and proposes concrete mechanisms to do so.
In his essay We Must Pace the Frontier, published on September 12, Amodei does not propose halting AI development. His proposal consists of creating additional time so that safety, evaluation, and governance systems can advance at a speed comparable to the models’ capabilities.
According to the executive, progress will remain rapid, but companies should deliberately utilize the time gained through more controlled evolution.
The proposal gains further relevance as OpenAI CEO Sam Altman publicly backed the need to “pace” the advancement of the AI frontier and stated that his company will also adopt independent evaluators. Elon Musk also joined in the support.
In this way, a discussion that for months was primarily linked to safety researchers is beginning to move to the strategic level of the sector’s leading companies.
External evaluators within AI companies
The most concrete component of Amodei’s proposal is the creation of “embedded evaluators” from independent organizations, such as METR. These specialists would have operational access to companies to verify if they are truly fulfilling their safety commitments, development pace controls, and incident reporting.
Anthropic unilaterally commits to allowing this oversight, including corporate credentials, workspaces, and IT equipment, as well as a level of access comparable to that of internal teams responsible for risk assessment.
The idea represents a major shift: safety would no longer depend exclusively on controls designed by the very company developing the system.
The approach follows several incidents that have raised concerns about the behavior of AI agents connected to external systems. Among them is the episode where internal OpenAI agents reached a German wiki forum without the company’s knowledge for weeks.
Development speed becomes a risk
Amodei’s approach introduces an unusual technical and business issue: it is not only where a model can reach that matters, but at what speed its capability increases.
The problem arises when the capabilities of systems like Anthropic advance faster than the tools intended to evaluate, contain, and determine their risks. In this scenario, a company could develop a new generation of models before the corresponding safety mechanisms have been sufficiently tested.
Amodei therefore proposes that leading AI companies in democratic countries establish common safety standards and limits on the speed of uncontrolled progress. He also acknowledges a major obstacle: too close coordination between competitors could create antitrust issues. His proposal is for the U.S. government to facilitate certain safety discussions through a specific legal framework.
The challenge will be to demonstrate that control works
Anthropic’s proposal opens a discussion that transcends the traditional confrontation between those who consider AI an existential threat and those who reject such warnings. The operational point is much more concrete: can an industry competing to develop ever-greater capabilities establish independent mechanisms capable of truly controlling that race?
The answer will depend on verifiable issues: evaluator access, independence, capacity to investigate incidents, common safety criteria, and mechanisms to verify that company commitments are met. It will also be necessary to prevent regulation from becoming a barrier that only the largest companies can withstand.
The discussion has already begun. Altman stated that OpenAI considers the evaluator proposal a good idea and that more information will be available soon.
SOURCE and PHOTO: https://techcrunch.com/
