OpenAI and Anthropic Want More Control Over What Independent AI Safety Reports Reveal

OpenAI and Anthropic publicly support independent AI safety evaluations, but a dispute over publication rights is complicating that promise. According to Reuters, both companies have discussed how much influence they should have over reports produced by outside evaluators. The issue is bigger than protecting confidential information. These reviews help outsiders understand whether increasingly powerful AI systems pose risks that companies might not disclose themselves.

OpenAI and Anthropic negotiate independent AI safety report rules

The news agency reported that both AI companies have sought significant influence over what external evaluators can publish. The discussions reportedly involve protecting trade secrets and preventing sensitive information from reaching competing AI labs.

Proposals have included allowing company employees to observe embedded evaluators and restricting evaluators from working inside rival labs simultaneously. However, technical information considered proprietary could also be essential for assessing an AI model’s safety.

Anthropic told Reuters that evaluators it already works with control their methods, criteria, conclusions, and publication decisions. For new evaluation firms, Anthropic said it reserves the right to comment on and approve reports while relationships are established.

Meanwhile, OpenAI said it does not seek approval rights over entire reports. The company said it redacts confidential information and allows evaluators to disclose whether major redactions occurred. These differences matter because the two companies’ public support for independent reviews does not automatically guarantee unrestricted publication.

AI safety transparency faces a bigger test

The debate follows growing pressure for stronger independent oversight of frontier AI models and agents. OpenAI and Anthropic have also discussed AI safety with Google DeepMind, according to earlier reports on industry safety talks. More than 200 researchers have reportedly backed calls for evaluator independence, broad access, and protection against retaliation. Without common standards or binding rules, AI companies could retain considerable influence over how their systems are assessed.