Anthropic and OpenAI are in negotiations over how much AI evaluators can publish, according to sources familiar with the talks. The discussions could set new limits on what independent researchers reveal about the companies' artificial intelligence systems, potentially curbing transparency in a field where outside scrutiny has become a key check on rapid development.
The push to control the narrative
Evaluators—often third-party researchers or red teams—test AI models for flaws, biases, and dangerous capabilities. Their public reports have become a primary way the public and policymakers learn what these systems can do. But the two leading AI firms are now negotiating the terms under which those findings see the light of day. The specifics remain undisclosed, but the talks center on how much detail evaluators can share without running afoul of contracts or competitive concerns.
It's a sensitive balance. Companies argue that unfettered disclosure can expose trade secrets or hand adversaries a roadmap for misuse. Critics counter that secrecy shields the same systems from the accountability that makes technology safer. Neither firm has commented publicly on the negotiations.
Why evaluators matter
Independent evaluation has been a cornerstone of AI safety efforts. When a model is found to generate harmful content, leak private data, or exhibit unexpected biases, evaluators document it. Those documents inform regulators, journalists, and the public. If evaluators can't publish freely, that feedback loop weakens. The negotiations between Anthropic and OpenAI could effectively standardize what gets published—or what gets buried.
The two companies are rivals, but they share a common interest in shaping how their products are perceived. Coordinating on evaluator disclosure rules would give them a unified front, but it also raises the risk of collusion concerns. No government body has announced involvement in the talks.
Accountability on the line
Public trust in AI hinges on the belief that someone outside the companies is checking the work. If evaluators are muzzled, that trust erodes. The negotiations come as lawmakers in multiple countries debate AI regulation, often relying on third-party audits to inform their decisions. A clampdown on evaluator publishing could leave those efforts flying blind.
It's not just about what evaluators find—it's about who gets to know. The outcome of these talks will shape the information ecosystem around two of the most influential AI developers. For now, the details are behind closed doors, and neither Anthropic nor OpenAI has indicated when a decision might be reached.
