The report
The October 30 IETF individual Internet Draft proposes a robots.txt policy group identified by User-Agent: __usage_ai-training. It reuses ordinary path-level Allow and Disallow syntax to express whether a publisher permits AI training on particular pages. The draft explicitly distinguishes training preference from permission to crawl or retrieve a page; a wildcard User-Agent: * does not automatically match the new purpose token.
The proposal is not an approved RFC, and robots text cannot stop a noncompliant crawler. It also has no bearing on whether an AI agent should trust a webpage's instructions. The training-versus-grounding concept links this policy question to the separate data-void search manipulation case, where retrievers used malicious pages as evidence. Publishers should not confuse a usage preference with technical protection against compromised content or unauthorized downstream actions.
Adoption is not the same as a deployed standard
The draft proposes a shared syntax; provider announcements concern future or staged support and may coexist with provider-specific preferences and private commercial rights agreements. A website owner therefore needs to inspect each provider's current implementation before assuming that publishing the special group changes live collection. The rule is scoped to AI training use and does not authenticate the publisher, permit reuse in a grounded answer or allow a browser agent to follow page instructions. Even a perfectly formed robots.txt cannot revoke data already obtained or compel a noncompliant crawler. This is an issue of declared use preferences, not a replacement for action authorization.
What the declaration actually says
The draft's example uses User-Agent: __usage_ai-training followed by Disallow: / and Allow: /press-kit/. Those lines express a preference about use of content for model training; they do not automatically alter whether a search crawler may retrieve the page, and the absence of a policy group is not an affirmative grant of rights. The document itself is an individual Internet-Draft, published October 30 and scheduled to expire May 3, 2027 if not advanced or revised. Claims about future vendor support should remain provider-specific. The practical distinction between training permission and retrieval access is separate again from the trustworthiness of material fetched for an answer.