Google, OpenAI, Anthropic and Microsoft have backed a shared signal that will let website owners express an AI training opt-out across participating providers, reducing the need for separate company-by-company instructions. The commitments announced on October 30 support a proposed robots.txt mechanism, with implementation schedules varying across the companies.
The move addresses a practical problem for publishers: communicating the same training preference to multiple AI developers through different controls. The proposal remains under consideration at the Internet Engineering Task Force rather than being a completed standard.
How the signal works
The mechanism uses a shared policy token, __usage_ai-training, in a website’s robots.txt file. Publishers can pair it with a Disallow rule to express a site-wide training opt-out, while Allow rules can make exceptions for specific directories, such as a press kit.
The distinction is between how content is used and whether it can be accessed: the training preference does not prevent crawling altogether or automatically remove pages from ordinary search results. Existing provider-specific controls will continue alongside the shared signal.
Google says that when a website opts out through the shared token without explicitly specifying Google-Extended, it plans to treat the instruction as a Google-Extended block. Publishers can still use Google’s own control to communicate a company-specific preference.
Rollout stretches into 2027
Anthropic has already begun supporting the directive, while Microsoft says it will start honouring the new preferences on February 1, 2027. Google plans support in early 2027, and OpenAI has pledged implementation in the coming months, according to Engadget.
That staggered rollout means the announcement does not create an immediately operational opt-out across all four companies. Publishers’ existing company-specific instructions remain relevant while implementation proceeds.
Microsoft separates grounding from training
Microsoft is also introducing separate controls for AI training and grounding—the use of web content to help generate AI answers. Its Limited Grounding option will allow content to support Microsoft’s own AI products while excluding its third-party grounding services.
The distinction would let a publisher refuse model training while still allowing its articles to support or be cited in an AI answer. Microsoft says these preferences will preserve eligibility for traditional Bing search results.
Temporary grounding controls are scheduled to become available on December 15, 2026, for eligible sites using older NoArchive or NoCache directives. Publishers relying on those directives have until March 1, 2027, to configure the new preferences if they want to maintain their restrictions.
A voluntary, prospective control
The shared mechanism remains a voluntary commitment by participating companies, not a technical barrier preventing all AI systems from accessing a website. Its practical reach depends on providers implementing and honouring the preference.
It also cannot undo past collection of website content. For publishers, the immediate change is a simpler way to communicate future training preferences—not a means of withdrawing material already gathered by AI developers.
