The shared robots.txt signal is an Internet-Draft, still short of a finished standard. On 30 October three companies said they plan to support it, and none has switched it on.
A website owner can state a preference against AI training in robots.txt, but none of the three companies that announced support honours it yet. The signal is a policy token, __usage_ai-training, written as a User-Agent group.
It comes from an Internet-Draft, draft-kelley-aipref-training-preference-signal-00, dated 30 October 2026. Its author is Kevin Kelley of Anthropic. It is an early proposal in the IETF AI Preferences working group.
What the draft says
The draft on the IETF archive lists its intended status as Standards Track and expires on 3 May 2027. It adds no new directive names and no new grammar. A publisher heads an ordinary robots.txt group with User-Agent: __usage_ai-training, for example followed by Disallow: / and Allow: /press-kit/, and the longer matching rule decides each path.
What the three companies said
On 30 October, OpenAI, Google and Microsoft announced plans to support it, and none said it has implemented it. OpenAI says it "will honor a common signal in robots.txt" and aims "to implement the signal in the coming months". Google says it aims "to support this new control along with other AI providers in 2027".
Microsoft will honour the token from 1 February 2027, per its Bing blog, which calls the work "significant progress that merits adoption".
OpenAI and Google both keep their existing controls in force. OpenAI says the new signal "will not replace" its existing robots.txt crawler controls. Google says existing controls "will continue to function alongside" it.
Google also plans to treat a token opt-out as a Google-Extended disallow when the site does not specify Google-Extended. That is the training control Google has offered since 2023.
Limits built into the draft
A wildcard group, User-Agent: *, "MUST NOT be interpreted as matching the policy token", so a blanket crawler rule does not express the training preference. A file with no group for the token expresses no preference through this mechanism.
The token "is not a directive for crawling behavior", and "private agreements MAY override preferences expressed via this mechanism". The preference covers training that modifies the learned parameters of a generative AI model. Other kinds of model training fall outside it.
The draft's acknowledgments name staff from Google, Microsoft, OpenAI, Anthropic, Meta, Cloudflare, News Corp and the BBC, among others. The draft lists them "for identification only". They are credited contributors, not a list of companies committed to honour the token.
What to watch
CTOL's view is that publishers can add the token now, but none of the three says it honours it today. Every start date is in the future, and existing provider controls still apply alongside the token.
Watch for Microsoft's start on 1 February 2027, OpenAI's implementation update and Google's 2027 support. Only Microsoft has given an exact date. The draft's expiry on 3 May 2027 is not a deadline for any of them. A published start would show that a company says it reads the token.
It would not show that any model has stopped training on a site's pages.