{"schema":"sanctity-signal/0.2","record_type":"signal","id":"aipref-training-signal-e067","domain":"governance","drop_id":"signal-e067","occurred_at":"2026-10-30","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Anthropic","action_type":"filing of an individual internet draft defining a single robots.txt policy token for expressing a preference against AI training, with 3 rival developers reported as backing it","observed":"The Internet Engineering Task Force datatracker carries draft-kelley-aipref-training-preference-signal-00, \"Machine-Readable Training Preference Signal\", by Kevin Kelley of Anthropic, dated October 30, 2026 and expiring May 3, 2027. We read it at datatracker.ietf.org. It is an individual internet draft with intended status Standards Track, no stream assigned, an IESG state of \"I-D Exists\", and no endorsement or formal standing from the task force. It defines the policy token __usage_ai-training. A site operator writes ordinary robots.txt groups whose User-Agent line names that token: a Disallow under it expresses a preference against AI training for the matched paths, an Allow expresses a preference in favour, and a file with no group naming the token expresses no preference. The draft adds no new directives or grammar, relying on the existing path matching, longest-match precedence and group structure of RFC 9309. Wildcard groups do not count as matching the token. The vocabulary follows the task force aipref working group's definitions of AI training and generative AI models. Preferences are independent of crawling rules: a crawler's own product-token groups keep their usual meaning, and the token is not itself a crawling directive, though consumers may use it to decide what to fetch. More specific rules take precedence over general ones, private agreements may override the preferences, and the draft has no IANA actions. Reported, not read by us at the companies' own posts, which we could not retrieve: SiliconANGLE, published October 30 and updated at 22:00 Eastern that day, says Google, Microsoft, Anthropic and OpenAI announced plans to support the standard; that Microsoft and OpenAI \"intend to add support for the crawling standard in the coming months\"; that Google will follow \"sometime next year\"; and that Meta has expressed support for the idea without committing to implement it. SiliconANGLE describes a Microsoft setting called AI Grounding, for cases where a model fetches a page to answer a question. Later summaries give Microsoft a February 1, 2027 start, which we could not confirm at Microsoft.","evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"IETF datatracker: draft-kelley-aipref-training-preference-signal, dated Oct 30, 2026, Kevin Kelley, Anthropic","url":"https://datatracker.ietf.org/doc/draft-kelley-aipref-training-preference-signal/"},{"role":"corrob","label":"SiliconANGLE: Google, Anthropic and others to support new AI crawling opt-out standard, Oct 30, 2026","url":"https://siliconangle.com/2026/10/30/google-anthropic-and-others-to-support-new-ai-crawling-opt-out-standard/"},{"role":"corrob","label":"Shopifreaks: Google joins Microsoft, OpenAI and Anthropic in backing a single robots.txt signal","url":"https://www.shopifreaks.com/google-joins-microsoft-openai-and-anthropic-in-backing-a-single-robots-txt-signal-that-lets-sites-opt-out-of-ai-training-in-2027/"},{"role":"corrob","label":"Dev Community summary naming the token and the draft, Oct 31, 2026","url":"https://dev.to/delimitter/openai-google-microsoft-and-anthropic-to-support-a-unified-ai-training-opt-out-token-in-robotstxt-4204"}]},"strength":{"cost":"Low","commitment":"Medium","reversibility":"High","persistence":"Months","rationale":"Cost is low, and we are not pretending otherwise. Posting an individual internet draft takes an afternoon, costs nothing and binds no one. Commitment is medium rather than low only because 4 companies that compete for the same training data are reported to have said in public which token they will read, and a published promise about a specific string is testable in a way a policy page is not. Reversibility is high: the draft expires on May 3, 2027 unless replaced, and a company that stops honouring the token breaks no contract. Persistence is months. This is scored as the modest action it is. What makes it worth a card is who converged, not how much it cost them."},"shift":{"prior":"Every AI company reads a different opt-out tag.","action":"The Internet Engineering Task Force datatracker carries draft-kelley-aipref-training-preference-signal-00, \"Machine-Readable Training Preference Signal\", by Kevin Kelley of Anthropic, dated October 30, 2026 and expiring May 3, 2027. We read it at datatracker.ietf.org. It is an individual internet draft with intended status Standards Track, no stream assigned, an IESG state of \"I-D Exists\", and no endorsement or formal standing from the task force. It defines the policy token __usage_ai-training. A site operator writes ordinary robots.txt groups whose User-Agent line names that token: a Disallow under it expresses a preference against AI training for the matched paths, an Allow expresses a preference in favour, and a file with no group naming the token expresses no preference. The draft adds no new directives or grammar, relying on the existing path matching, longest-match precedence and group structure of RFC 9309. Wildcard groups do not count as matching the token. The vocabulary follows the task force aipref working group's definitions of AI training and generative AI models. Preferences are independent of crawling rules: a crawler's own product-token groups keep their usual meaning, and the token is not itself a crawling directive, though consumers may use it to decide what to fetch. More specific rules take precedence over general ones, private agreements may override the preferences, and the draft has no IANA actions. Reported, not read by us at the companies' own posts, which we could not retrieve: SiliconANGLE, published October 30 and updated at 22:00 Eastern that day, says Google, Microsoft, Anthropic and OpenAI announced plans to support the standard; that Microsoft and OpenAI \"intend to add support for the crawling standard in the coming months\"; that Google will follow \"sometime next year\"; and that Meta has expressed support for the idea without committing to implement it. SiliconANGLE describes a Microsoft setting called AI Grounding, for cases where a model fetches a page to answer a question. Later summaries give Microsoft a February 1, 2027 start, which we could not confirm at Microsoft.","update":"One token may cover the 4 largest, from 2027."},"inference":{"established":["An individual internet draft dated October 30 defines the token.","The token signals a training preference only, not crawling or search.","The draft has no task force standing and expires May 3, 2027."],"unknown":["That anyone honours it. Compliance is voluntary and unaudited.","The start dates. No company post we could read gives one.","That Meta joins. It is reported to back the idea without committing."]},"narrative":{"hook":"Anthropic filed an internet specification for a single AI training opt-out signal.","plain":"An Anthropic engineer filed an internet draft defining one robots.txt token for refusing AI training. Reported: Google, Microsoft and OpenAI said they will honour it, on timelines running into 2027.","story":["A publisher today writes a separate rule for each crawler and still cannot separate training from search. One token, honoured by the largest developers, turns refusal into something a court or a regulator can check. It remains a preference, not a lock. The dates each company gave are the test."],"trail":[{"d":"SEP 4","t":"OpenAI, Microsoft and the news publishers all ask one judge to decide the largest AI copyright case without a trial. Edition 013's court card."},{"d":"OCT 1","t":"A federal judge throws out the publishers' antitrust case against Google's AI summaries, leaving copyright as their remaining route. Edition 039's court card."},{"d":"OCT 29","t":"The Wikimedia Foundation moves to file an amicus brief in the consolidated OpenAI copyright case, arguing training is likely fair use and retrieval is not. Recorded in Undercurrents."},{"d":"OCT 30","t":"The draft defining __usage_ai-training is posted at the standards body. Reported the same day: 3 rival developers say they will honour the token, and Meta does not commit."},{"d":"OCT 30","t":"Britain's competition regulator opens an abuse case over Meta's terms for taking publisher content for AI features. This morning's second card."},{"d":"Reacts","r":true,"t":"News publishers and the trade bodies that have asked for one tag for years; Cloudflare and the other firms selling crawl controls; Meta, the one large developer not committed; the task force aipref working group, which has not adopted this draft; and every site owner who now has a text file to edit."}],"glossary":[{"term":"robots.txt","def":"A text file at a website's root telling automated visitors which paths they may fetch. It carries no enforcement; compliance is a convention."},{"term":"Internet draft","def":"A working document at the internet standards body. An individual draft carries no endorsement and expires in about 6 months unless replaced."},{"term":"Product token","def":"The name a crawler answers to in robots.txt, such as a search engine's bot name. Rules are written per token."},{"term":"Grounding","def":"Fetching a live page to answer a question rather than to train on it. Same content, different use, and until now the same tag."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Refusal is becoming machine readable","text":"One checkable token is what enforcement and licensing both need first.","abstain":false},{"id":"b","label":"Narrow technical housekeeping","text":"A tag in a text file, honoured when convenient, changes very little.","abstain":false},{"id":"c","label":"A deliberate move before regulation","text":"A voluntary standard filed now is the alternative to a statutory one.","abstain":false},{"id":"d","label":"Normal standards work","text":"Crawler conventions have always been written this way, by one author.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"Two or more of Google, Microsoft, OpenAI and Anthropic publish documentation stating they honour __usage_ai-training, or the aipref working group adopts the draft, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"The draft expires on May 3, 2027 with no working group adoption and no company documentation naming the token, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"c","direction":"strengthens","observable_event":"A regulator, court or legislature cites the token or an equivalent machine-readable training preference in guidance, an order or bill text, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"}]}