Anthropic and OpenAI are reportedly seeking control over evaluator publications, according to sources familiar with the matter. The two leading AI companies want a say in what independent evaluators can publish about their models. The move has sparked concerns that it could undermine transparency and trust in the AI industry, and it raises thorny questions about trade secrets.
The push for control over evaluator publications
Evaluators—often third-party researchers or organizations—test AI systems for safety, bias, and performance. Their findings are typically published openly, allowing the public, regulators, and other companies to assess the technology. But Anthropic and OpenAI are now reportedly angling to have more control over what gets released. The exact mechanism isn't clear: it could involve contracts, review processes, or restrictions on publishing certain details. Both companies have emphasized the need to protect trade secrets, arguing that unfettered disclosure could hand competitors an advantage or reveal sensitive methods. Yet the same secrecy could also hide flaws.
Why transparency matters for AI trust and innovation
Transparency has been a cornerstone of AI research. When evaluators publish their methods and results, others can reproduce tests, spot weaknesses, and build on successes. That cycle drives innovation and builds public confidence. If companies can veto or shape evaluator publications, that cycle breaks down. Independent researchers might hesitate to flag problems if they fear pushback. Users might not learn about risks until it's too late. And regulators, already struggling to keep pace with AI, would have less material to work with. The affected parties aren't just the companies and evaluators; they include anyone who relies on AI systems—which is increasingly everyone.
The trade secret dilemma
There's a real tension here. AI models are expensive to build, and companies have legitimate interests in protecting proprietary techniques. Trade secret law allows businesses to keep certain information confidential as long as they take reasonable steps to do so. But applying that framework to evaluator publications is tricky. Evaluators need enough access to do their jobs, and the public needs enough information to trust the results. Overly broad secrecy claims could stifle the very scrutiny that makes AI safer. Anthropic and OpenAI haven't publicly detailed their proposals, so it's unclear where they'd draw the line. What counts as a trade secret versus a public safety concern? That question is likely to be fought over in contracts, courts, and policy debates.
What's at stake for the AI industry
The AI industry is watching closely. If Anthropic and OpenAI succeed in controlling evaluator publications, other companies may follow. That could set a precedent for how AI firms interact with independent testers. Startups and smaller players might feel pressured to adopt similar policies to stay competitive. Meanwhile, academic and nonprofit evaluators could find their work constrained. Some might refuse to sign restrictive agreements, reducing the pool of independent expertise. Others might comply but self-censor. Either way, the quality and quantity of public evaluations could suffer. The outcome will shape not only what we know about AI systems but also who gets to decide what's safe.
No clear resolution yet
Neither Anthropic nor OpenAI has announced a formal policy or timeline. The discussions are reportedly ongoing, and no agreements have been made public. For now, the situation remains fluid. Evaluators, researchers, and watchdog groups are likely to push back if concrete restrictions emerge. Regulators in the US and abroad, already drafting AI rules, may take an interest. The core conflict—between protecting trade secrets and preserving transparency—won't be easy to resolve. How Anthropic and OpenAI proceed will be a test case for the industry. And for anyone who uses AI, the stakes are high: less transparency could mean more risk, unseen.
