IETF ai-control debates split AI inference preference categories
A draft would let rights holders treat autonomous system use of assets differently from content a human user supplies at inference time.
Nate Hake has put forward two new preference categories for the IETF ai-control vocabulary, splitting AI inference according to whether a system or a human user identifies the asset. Developed with Paul Keller after the Vienna meeting, the draft is meant to sit cleanly beside a training category so rights holders can cover the full range of model uses, and so providers can respond with more than a single yes or no.
"AI System Inference" would cover use of an asset to generate synthetic content when the system selects that asset. "AI User Input" would apply when a human specifically identifies the asset by providing it or a location from which it can be retrieved. Inference, in both cases, means any use beyond changing a model's learned parameters. Hake argued the pair leaves no undefined gap and lets declaring parties distinguish autonomous bulk processing from end-user work on individual assets.
Keller stressed the consumer side of that design. A single inference category, he wrote, admits only honouring the preference for all inference or for none. Two categories open a third path: honour the preference when the system fetches the asset, and not when the user supplies it. That choice would also be legible to the rights holder. Leonard Rosenthol of Adobe said the split matches shipping practice in Photoshop Generative Fill, which already checks Content Credentials on a user-supplied reference image and can warn before proceeding while still allowing the user to override.
Mirja Kuehlewind challenged whether anyone else can find the line. She questioned how much indirection still counts as user identification: a pasted link, a description, or a prompt that tells an assistant to pull similar material from a named site. Interfaces and agent behaviour will keep shifting, she said, so a blurry boundary produces outcomes users cannot understand. Chris Needham agreed the distinction is real but needs sharper wording, warning that phrases about a user-specified location could be read to include open-ended find-and-summarise prompts.
Hake held that the test is whether a human specifically identified the asset, typically by upload or direct link, and that any vocabulary will have edge cases receivers must resolve. Several participants favoured settling the wording in London rather than continuing definitional sparring on the list.