Anthropic and OpenAI reportedly want the final say over what independent evaluators can make public about their models.
Reuters reports, citing two people familiar with the discussions, that both companies have pushed outside reviewers to accept limits on what they publish, citing worries that findings could expose trade secrets. The push comes just weeks after both labs pledged to give evaluators a free hand to publish their results. Anthropic CEO Dario Amodei has publicly backed independent scrutiny of AI systems. Reuters did not name the evaluators involved or specify exactly which disclosures the labs want to restrict.
The problem is simple: evaluation only has teeth if the evaluator can say what they found, not just what the company is comfortable with. A reviewer who needs sign-off before publishing isn't independent; they're a contractor with a non-disclosure agreement. That distinction matters more for AI labs than most industries, since the main evidence the public gets about model safety and capability often comes from these outside tests.
Promising transparency and then negotiating the terms of that transparency answers its own question about how much transparency was really on offer.