IETF AI preference draft splits training from use
Revision 08 of the IETF’s AI Preference Vocabulary draft, published this week, restructures its usage categories: the previous single “AI Model Training” section splits into separate “AI Training” and “AI Use” sections, while a distinct “Search” section — covering how AI systems may use content to answer search queries — is retained and renumbered. The draft, co-authored by Google’s Martin Thomson and Paul Keller, defines an AI model for the vocabulary’s purposes as one used to “generate synthetic content,” explicitly excluding models used only for classification, ranking, or scoring. The new revision is set to expire March 18, 2027 unless renewed.
Why it matters: A standard splitting "AI training" from "AI use" from "search" gives sites a more precise vocabulary to say yes to being indexed and cited while saying no to being used as training data — the distinction publishers blocking crawlers have been drawing informally, here formalized in draft.
The record: llms.txt
Via IETF AI preferences vocabulary draft ↗
Posted to the wire September 16, 2026. Edited by Joe Balewski.