Google, Microsoft, OpenAI and Anthropic back one AI training opt-out

On October 30, Google, Microsoft, OpenAI and Anthropic each said they plan to honor one shared signal that lets a website refuse AI training. Today an owner has to name each company’s crawler. The new signal is a single entry in robots.txt, the file that already tells bots where they may go.

The entry is a “policy token” called __usage_ai-training. A site adds a rule for it, such as disallowing everything, and the participating companies read that as a no to training. Google says that if a site uses the token and has not set Google-Extended, it plans to treat the token like a Google-Extended block. Microsoft also adds page-level settings for training and a separate one for “grounding,” meaning answers that pull from live pages.

Timing differs. Anthropic says it has started supporting the signal. OpenAI says it aims to in the coming months. Microsoft says its new preferences start taking effect February 1, 2027, and Google says it aims for 2027.

The catch: This is a request, not a lock. The signal comes from an early draft with no formal IETF standing, it covers training rather than crawling, and it doesn’t pull back content already collected. Microsoft also warns that blocking grounding can sharply cut how visible your content is in AI answers.

What to do now

• Open your robots.txt (usually yoursite.com/robots.txt) and see which AI crawlers you already block.
• Decide whether you want training allowed, blocked or split by section of the site.
• If you opt out, add the token rule after the companies support it; Anthropic is the only one that says it has started.
• Don’t block grounding unless you accept less visibility in AI answers.
• Ask your web host or site builder whether it lets you edit robots.txt.