A new shared signal for AI training: the IETF draft “Machine‑Readable Training Preference Signal”

On 30 October 2026 the individual Internet‑Draft “Machine‑Readable Training Preference Signal” (draft‑kelley‑aipref‑training‑preference‑signal‑00), written by Kevin Kelley of Anthropic, was published at the IETF. The 10‑page document introduces the policy token __usage_ai‑training, which goes in the User‑Agent line of an ordinary robots.txt group to express an AI training preference path by path. The example given is “User‑Agent: __usage_ai‑training / Disallow: / / Allow: /press‑kit/”: no training anywhere on the site except the /press‑kit/ directory.

The draft leaves robots.txt’s path syntax, longest‑match rule and group structure untouched, adding no new directives and no extra grammar. It also specifies that wildcard groups (User‑Agent: *) must NOT be read as applying to the token; if no group mentions __usage_ai‑training, the site has expressed no preference. The signal covers only the use of content to change the parameters of a generative model, not crawling behaviour (section 4.4); to limit crawling, the individual crawlers’ tokens still apply. The document states that it is not endorsed by the IETF and has no formal status in the standards process; it expires on 3 May 2027 unless revised.

The same documentation points out that Google‑Extended is a standalone product token and does not affect a site’s inclusion in Google Search. Google says it intends to support the AIPREF signal from early 2027, keeping the Google‑Extended token separate, which has no bearing on inclusion in search results.

OpenAI announced on 30 October 2026 that it will honour the signal and plans to implement it in the coming months, without replacing its crawlers’ existing controls; the work builds on that of the IETF AI Preferences group, and OpenAI is inviting other providers to adopt it.

Operational unknowns and documented gaps - The draft is still an individual document and may change before any formal adoption. - The signal is a voluntary preference: it is not an access‑control mechanism, and non‑cooperating crawlers may ignore it. - No primary source confirms that the signal extends to content already collected or to models already trained. - Specific adoption dates for other providers are not backed by primary sources.

If adopted, the token would let sites express a single preference valid for every participating provider, separating training from crawling, instead of a different token for each provider (GPTBot, Google‑Extended, ClaudeBot, etc.).

— Pixie