The companies backed a shared standard that lets websites signal they don’t want their content used to train AI models.

On October 30, 2026, OpenAI, Google, Microsoft and Anthropic backed a shared website signal for opting out of AI training. The proposed IETF draft, “Machine-Readable Training Preference Signal,” by Anthropic standards lead Kevin Kelley, uses the robots.txt token `__usage_ai-training`; a site can pair it with `Disallow: /` to express a site-wide preference. Anthropic says it has already begun honoring the signal; OpenAI plans implementation in the coming months, Microsoft from February 1, 2027, and Google sometime in 2027. The draft expires May 3, 2027, and adoption is voluntary. For publishers, one instruction could replace separate training opt-outs for participating crawlers, easing a fragmented process amid disputes over the use of news and other web content. But the token concerns training, not whether a bot may fetch a page, and it does not undo training already completed. Microsoft says sites remain eligible for Bing search results; it also plans separate controls for AI grounding, or using sources in generated answers.