What happened
On October 30, 2026, Google, Microsoft and OpenAI each announced that they would honor a shared robots.txt instruction for AI training. That same day, Anthropic engineer Kevin Kelley told the IETF AI Preferences working group (AIPREF) mailing list that Anthropic had started implementing support for his Internet-Draft, draft-kelley-aipref-training-preference-signal-00. The draft names Kelley as its sole author, lists an intended Standards Track status, and expires May 3, 2027; it is not a formally adopted IETF standard.
The proposed mechanism uses an ordinary robots.txt group whose User-Agent field names the policy token rather than a crawler. The draft's example is User-Agent: __usage_ai-training, followed by path rules such as Disallow: / and Allow: /press-kit/. It uses the existing path-matching and longest-match precedence of RFC 9309, the Robots Exclusion Protocol. A site can therefore express a preference by path without creating a separate set of instructions for every provider.
The draft says it adds no directive names or grammar to robots.txt; parsers that do not implement the token should behave as before. Consumers must match policy tokens exactly apart from case. A wildcard User-agent: * group does not match the policy token, so an existing blanket bot rule does not by itself express a shared training preference. Nor does a block on a provider-specific token such as Google-Extended count as the shared preference for other providers.
The announcements give different support dates. Microsoft says it will honor the token from February 1, 2027, and is retiring its NoArchive and NoCache tags for its AI opt-out; sites relying on those tags must switch by March 1, 2027. Google says it aims to support the signal in early 2027. OpenAI says implementation is planned in the coming months. Anthropic says implementation has begun but gives no live date. The draft explicitly distinguishes use from access: the token communicates a preference about training, not a crawling instruction. Google says it will also apply to Gemini grounding, while search rankings are unaffected.
What the archive adds
This is a new implementation update to a story covered in Making it easier for websites to express AI training preferences, which described OpenAI's stated intent to honor a shared signal. The archive also has Google, Microsoft, OpenAI and Anthropic back a shared robots.txt policy token for opting out of AI training, which covers the same event; this source adds provider commitments, timing, and operational caveats.
Trend signal
This supports Machine readers need operator identity: providers are trying to make automated use legible through a machine-readable policy token rather than relying only on crawler names. The signal remains a voluntary preference with staged support, not a universal enforcement or identity system.
Why now
The announcements and draft appeared together on October 30, 2026. The source describes the shared mechanism as a response to site operators having to maintain separate crawler rules for different AI developers. Microsoft has also set a transition deadline for sites currently relying on its NoArchive or NoCache opt-out.
Connections to other events
A previous Telegram post described Cloudflare's Disallow AI Training control as separating permission for search from permission for training. The shared token addresses a related distinction at the robots.txt level: it communicates a training preference while leaving crawling and search ranking separate. The source also connects the mechanism to RFC 9309, whose existing path rules it reuses.
Why it matters
For publishers, a shared token could reduce the upkeep of provider-by-provider training rules. For teams that collect web data, it creates a clear implementation task: parse the explicitly named policy group and preserve its path-specific meaning, rather than treating every wildcard or vendor-specific block as equivalent. That distinction is important for both compliance and data provenance.
The practical effect is not immediate or uniform. Microsoft gives a firm start date and a migration deadline for its older tags; Google, OpenAI and Anthropic describe different, less precise timelines. Until a provider implements the draft, adding the token alone cannot be assumed to change that provider's behavior. Teams should also keep access controls and training-use preferences separate: the draft says the token is not a directive for crawling.
Limits and uncertainty
The source reports company announcements and an Internet-Draft, not a finalized standard or evidence that every provider has deployed support. OpenAI and Google have not supplied exact dates in the material, and Anthropic's statement that implementation has started does not establish when it will be live. The token records a preference; the source does not establish a technical enforcement or auditing mechanism. Nor does it say that the token changes permissions for uses other than AI training, except Google's stated inclusion of Gemini grounding. Existing robots rules and provider-specific signals retain their prior meanings.
Reader takeaway
The shared __usage_ai-training token is a potentially simpler way to express a site-wide or path-specific training preference across participating AI providers. It is not a crawl block, does not count when placed only in a wildcard group, and will matter only as providers implement it. Publishers should check Microsoft's migration deadline and avoid assuming one robots rule already has universal effect.