The other half of the AI “frontier”
Anthropic’s CEO wants evaluators embedded in AI companies, with the same level of access as employees. Brussels can already require external evaluations. What it lacks is anyone to carry them out.
On 12 September, the head of Anthropic published the essay “We must pace the frontier”, in which he argues that the leading artificial intelligence (AI) labs have to slow the pace at which they develop new models. Dario Amodei wants to embed external evaluators in these companies, with a desk, a pass and a work computer, like any internal employee, and the freedom to publish everything they find. Shortly afterwards, Sam Altman, head of OpenAI, said it was a good idea and promised to do the same.
Over one weekend, the heads of the two American companies that sell access to the world’s most advanced AI technology publicly acknowledged that their own word on the safety of their models is no longer enough: someone from outside needs to check and confirm it. Amodei’s proposal, endorsed by Altman, is not a moratorium. The Anthropic co-founder says slowing down “does not mean stopping model training or technical progress”. So there is no deadline, but there is an underlying question: who gets to look inside a model before it is released?
AI translation from the Portuguese. First published in ECO on 16 September 2026.