Anthropic and OpenAI haggle over what outside AI auditors can publish

The trouble is that what a lab calls crucial intellectual property can overlap with what you need to judge a model's risk. A separate person with knowledge of past dealings said that overlap leads to arguments and redactions. Reuters also saw a contract whose language can bar evaluation firms from publishing confidential information without the company's consent.

What the companies say

Anthropic told Reuters that the description of seeking significant control does not fit all its work with independent evaluators. It says evaluators it has worked with before control their methods, criteria, conclusions and what to publish. For new firms, it reserves the right to comment on and approve reports while a working relationship is built.

OpenAI says it is not seeking approval rights over the full text of reports. Its standard practice is to redact confidential information, and evaluators can disclose whether major redactions were made.

Redaction versus control of conclusions

Whether the public and private positions truly clash turns on a distinction. Blacking out a trade secret is one thing. Approving or reshaping a report's findings is another. The companies' statements, as Reuters relays them, describe the first, while the two sources describe something closer to the second. Reuters does not spell out what controls beyond redaction the labs want, and we have not independently verified either account. The practical test is whether evaluators can tell readers that something major was cut, which OpenAI says they can.

The bigger problem: who picks the referee

Experts note that the labs choose who does the reviews. Researchers quoted by TechCrunch worried that companies could end up "shopping for evaluators" that aren't qualified or aren't interested in the most concerning risks. Without agreed standards or legislation, some say the companies could dictate how their own systems are examined. Both are also seen as heading toward possible major IPOs at trillion-dollar valuations, which raises the stakes on secrecy.

Anthropic said last month that Faculty, Accenture's specialist AI business, would be its first embedded evaluator, with access and reporting details still being worked out. Faculty CEO Marc Warner told Reuters that independence and transparency are critical if people are to trust evaluators.

What's still unknown: the final terms, and whether any redaction rules will be public. Watch for the Faculty arrangement details.

Should a lab ever get a say in a safety report about its own model?