Anthropic and OpenAI want to decide what outside evaluators can publish about their models, citing trade secrets and competitive risk.
Anthropic and OpenAI are negotiating how much outside evaluators can publish about their AI models. The goal is to prevent revealing trade secrets while keeping the public informed about safety and bias findings. Some contracts reportedly allow sharing confidential details without company approval, which raises concerns about independence. As the AI safety ecosystem grows, expect more talks and clearer rules on what evaluators can disclose and how they’re overseen. Next steps include finalizing publication rules and ensuring evaluators remain credible and independent.
Why it matters to you: independent reviews help you understand AI risks, but you don’t want secret methods leaking that could be misused. Watch for upcoming policy benchmarks and any new disclosure standards from labs and regulators.
What to do now: stay informed about any public evaluation reports from major labs, and look for official statements detailing what kind of findings will be published.
The news: Anthropic and OpenAI are pushing to keep tighter control over what outside evaluators can publish about their AI models. Reuters reported that the discussions center on preventing the disclosure of trade secrets or sensitive methods. The talks reflect a broader tension in the AI safety world: how to balance independent scrutiny with protecting confidential information that could affect a company’s competitive edge.
Why it matters: Independent evaluators help reveal how AI systems behave, where they fail, and what risks they pose. If labs curb what can be published, it can slow down public understanding of AI safety and bias. But labs argue that unfiltered results could expose strategic techniques or proprietary know-how. The current conversations show how much leverage labs may have when there are no strong, universal rules governing disclosure and review in AI development.
The discussions reportedly involve how much internal staff should oversee or veto what external evaluators publish, and whether evaluators who work at more than one lab can be restricted from sharing sensitive details. Some contracts hint at allowing publication of confidential information without the company’s consent, which would raise alarms for researchers and policymakers who prize transparency.
Anthropic says its approach varies by project, noting that teams it has worked with decide their methods, findings, and what to publish, while newer partnerships may retain the right to review and sign off on reports. OpenAI says it doesn’t want to approve entire reports; instead, it removes secret details and allows evaluators to flag when substantial edits were made. The firms stress they want independence for evaluators, but also want to prevent sensitive capabilities from leaking out.
As AI models become more capable and the potential impacts more widespread, labs are navigating a patchwork of contracts, safety reviews, and evolving expectations from developers, regulators, and the public. Without clear, shared rules, firms may use access as leverage or gatekeeping power at a critical moment in AI governance.
Transparency about how AI works matters because it shapes trust and safety. If outside reviews are too tightly controlled, you might not hear about certain risks or biases in a timely way. On the other hand, unrestricted publication could expose tactics that could be misused. The goal is to keep the public informed while safeguarding sensitive methods that keep products competitive and secure.
Watch for more details as negotiations continue. Expect clarifications on what kinds of findings must be shared publicly, what can be kept confidential, and how to verify that outside evaluators remain independent and credible.
Independent researchers or firms contracted to test and review an AI model’s safety, bias, and reliability.
Public findings can reveal trade secrets or sensitive development techniques, potentially harming competitiveness or security.
Announcements about new review frameworks, concrete rules on what can be published, and any shifts in who pays for and approves evaluator reports.
You're reading the quick version.
Read next
UK could lead global AI safety rules, Microsoft’s president warns
Microsoft president says AI should face the same kind of safety oversight as milk and washing machines, with incident reporting and independent evaluators before launch.
A pro-AI political network funnels millions through mystery Nevada LLCs to influence midterms. Here’s what readers in the US, UK and Canada need to know.
European Commission president Ursula von der Leyen floated a plan to let AI help officials reuse company data, so firms file once instead of multiple times.
The IMF flags high AI valuations and rising debt as factors that could trigger a broad market sell-off, with spillovers across the globe.
The Daily Brief
Every morning, the news that matters from AI, phones, apps and the people shaping tech. Explained in plain English. Free.
One email a day. No spam, unsubscribe anytime. Privacy policy
