Anthropic official brand assets (anthropic.com)
Anthropic and OpenAI negotiate how much AI evaluators can publish
Weeks after pledging editorial independence for outside testers, both labs are reportedly pushing for control over findings to protect trade secrets
Anthropic and OpenAI are negotiating with third-party AI evaluators over terms that could give the companies substantial control over which findings see daylight. That is according to a Reuters report published on October 13, 2026.
From pledge to fine print
On September 12, 2026, Anthropic CEO Dario Amodei committed to giving outside evaluators “ongoing, employee-like access” to the company. He also promised they would be free to publish findings on critical issues, including risks and safety practices.
There was one caveat. Amodei left room for necessary redactions, the polite term for blacking out the parts a company would rather competitors not read.
OpenAI’s Sam Altman fell in line with similar principles in mid-September 2026.
The negotiations have since tilted the other way. According to the research summarizing the Reuters findings, talks have swung toward the companies asserting editorial control over what evaluators release.
Both firms reportedly started by seeking strict oversight over which publications would be accepted. They then relaxed those constraints slightly in the days that followed.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Why embedded evaluators complicate everything
Much of the friction centers on evaluators who work embedded inside the labs rather than testing models from the outside.
Anthropic named Faculty, which the research describes as Accenture’s AI division, as its first embedded evaluator in September 2026.
Several specific conditions are reportedly on the table. One involves lab staff observing embedded evaluators while they work. Another involves restrictive clauses in existing contracts.
Those contracts may prevent evaluators from disclosing information without the companies’ consent. If that language survives, the “freedom to publish” pledge would come with a permission slip attached.
What this means for AI oversight
Several details are worth watching as talks continue. The first is how broadly redaction rights get defined in final agreements. The second is whether lab staff end up observing embedded evaluators as a standard condition.
The third is whether the restrictive disclosure clauses in existing contracts get rewritten or simply carried forward. The fourth is how Faculty’s arrangement with Anthropic plays out in practice, since it is the first embedded relationship on record.