The most important safety evaluations in the history of artificial intelligence are being conducted in secret. Not because the results are classified. Not because publication would create security risks. Because the labs conducting them do not want to be held accountable for what they find.
This is not a new observation. Researchers have been making it for years. What is new is the scale of the problem. As frontier models become more capable, the safety evaluations become more consequential — and the incentives for secrecy become stronger. We are approaching a point where the most important scientific work in the world is being done entirely behind closed doors.
The Self-Correction Problem
Science works through transparency. Results are published, methods are scrutinized, findings are replicated or refuted. This process is slow and imperfect, but it is the only mechanism we have for catching errors before they compound. When results are not published, errors do not get caught. They accumulate.
We are building the most consequential technology in human history using a methodology that systematically prevents the scientific community from checking our work.
The frontier AI labs have, in effect, opted out of this mechanism. They conduct internal safety evaluations, reach internal conclusions, and deploy systems based on those conclusions. External researchers cannot verify the evaluations. Regulators cannot audit the conclusions. The public cannot assess the risks. We are building the most consequential technology in human history using a methodology that systematically prevents the scientific community from checking our work.
The labs will tell you that they share results with trusted third parties, that they participate in red-teaming exercises, that they publish research on safety techniques. All of this is true. None of it is sufficient. Selective disclosure to trusted partners is not transparency. It is the appearance of transparency in service of continued secrecy.
What Radical Transparency Would Look Like
I am not arguing for the publication of model weights or training data. The security concerns there are real. I am arguing for the mandatory publication of safety evaluation results — the findings of the internal red teams, the capability assessments, the alignment evaluations — before deployment, not after.
This would be uncomfortable. Some of those evaluations would show things that labs would prefer not to show. That is precisely the point. The discomfort of transparency is the mechanism by which accountability works. A lab that knows its safety evaluations will be published has different incentives than a lab that knows they will not.
The objection that transparency would harm competitive position is real but not decisive. Yes, publishing safety evaluations would reveal information that competitors could use. The alternative — a race to deploy increasingly capable systems with no external accountability — is worse. The competitive cost of transparency is a price worth paying. The cost of continued secrecy may be civilization-scale.