Anthropic has announced it is installing embedded evaluators with employee-level access, tasked with checking its safety practices and reporting incidents. According to Bloomberg, Sam Altman has committed to placing third-party evaluators at OpenAI as well. This is exactly what we have been calling for since August.
In every article on safety, we have written the same sentence: a commitment that nobody can verify is not a commitment. Here is the first serious response to that objection.
Why this arrangement changes the nature of the problem
Let us recall the structural limit we highlighted regarding the slowdown announced in August: a company that sets its own rules, applies them and verifies that it is complying with them occupies all three roles. That does not produce verification, it produces declaration.
An outside evaluator with internal access breaks that configuration. They see what is happening rather than what they are shown.
The detail that matters is employee-level access. An auditor who receives documents prepared for them audits nothing. An auditor with the same access as an engineer can establish things for themselves — and above all, establish what they have not been shown.
The four questions that will decide its value
Who appoints them? An evaluator chosen and paid by the company they evaluate is in an uncomfortable position. This is the classic audit problem, familiar since the accounting scandals of the 2000s. The usual solution involves mandatory rotation and appointment by a third party.
What do they publish? A report submitted to management and a report made public do not serve the same function. The first informs, the second constrains.
What happens in the event of disagreement? If an evaluator believes a model should not be deployed and the company deploys it, what is the recourse? Without an answer to that question, the arrangement is advisory.
Are they protected? This is the most important point. We wrote it regarding the inventory requested by Congress: if reporting exposes you, nobody reports. An embedded evaluator must be able to speak out without risking their contract.
This arrangement exists elsewhere and it works, under precise conditions. In civil nuclear power, inspectors have permanent access to facilities, publish their findings, and their supervisory authority can order a shutdown. In aviation, reporting schemes protect the person who speaks up, which has brought accidents down far more than sanctions have. What these systems have in common is not the presence of inspectors: it is that they answer to an authority outside the company and that they can block.
What this adds
This announcement completes a rapid sequence. A British public institute observed deceptive behaviour from the outside. A lab published denominators that nobody possessed. And two labs are now opening up to internal observers.
Three weeks ago we wrote that trust in a technology we do not understand rests on four conditions: that someone understands it, that a record exists, that failures are analysed publicly, and that someone answers. The third has just made clear progress.
What to take away
This is good news, and it should be said without reservation before adding the caveats.
This mechanism is what was missing from all the announcements of recent months. It turns a declaration into an observation, and it does so voluntarily, without any law requiring it.
Its value will be measured against a single criterion: what will happen the first time an evaluator says no? If their opinion is published and acted upon, this mechanism will become a standard across the sector. If it is absorbed without consequence, it will join the list of commitments that reassure without constraining. We will know in the coming months, and that is exactly what will need watching.