Summary

  • RFC 9969 is an informational IAB workshop report: it records discussion and possible work, not a deployed AI-control protocol or a consensus mandate for entities’ views.
  • A preference must remain distinct from the attachment that carries it, the crawler that retrieves content, a later data-use decision, compliance evidence and any remedy.

The easy mistake in AI-content control is to call a published preference a control. RFC 9969 is unusually useful because it does not make that mistake. It documents the AI-CONTROL workshop’s discussion of preference signals, their attachment to Internet content and the points at which those signals lose force or evidence. The document is an IAB informational workshop report. It says the recorded views need not be IAB views, follows presentations and discussion notes without interpretation or validation, and does not attempt to capture a consensus. That status is not an embarrassment to edit away. It is the first operational boundary.

The report’s central timing problem is also an evidence problem. A publisher’s preference may be visible when a crawler encounters a resource. The data can then take paths that the location-bound statement does not accompany: a crawler may have multiple uses, a third party may collect the data, or a model may later use a corpus at inference time. RFC 9969 observes that a general-purpose foundation model generally cannot selectively apply a particular source’s preference while it performs a task. A crawl-time directive is therefore evidence of a signal at one moment, not proof of a later model-use decision.

Attachment changes what can be observed, but it does not erase this gap. The report discusses location attachment such as robots.txt, embedded metadata and registries. A site-level file can be too coarse for a multi-author or large service and does not automatically travel when content moves. Embedded preferences may stay closer to a work, yet can be stripped and are unsuitable for some content or mixed-producer pages. A registry can decouple a fingerprint from a location, but the report records serious Internet-scale deployment difficulties. These are different mechanisms with different failure modes; none is a certificate that a later use respected an owner’s preference.

The proposed starting points for technical work are correspondingly narrow: attach preferences by location, attach them in IETF-controlled formats, define a common core vocabulary, and define how multiple discovered preferences combine. The report explicitly leaves enforcement, registry-based solutions, identification or authentication of crawlers and content owners, and audit or transparency mechanisms out of initial scope. That list matters more than a hopeful reading of the word “control”. A vocabulary says what a signal purports to express. A combination rule can say how signals relate.

Neither answers who received the signal, which copy they used, what purpose they selected, or how anyone can later inspect the decision.

RFC 9309 provides the Robots Exclusion Protocol as a familiar attachment point, not a bridge over those missing records. A robots.txt observation can establish that a declared file was available at a location at a time. It does not establish that a particular crawler had the claimed identity, that a vendor’s internal collection was attributed correctly, that a dataset exclusion propagated, or that an inference result drew on or excluded a source. Treating it as any of those things would turn a narrow retrieval-layer fact into a conclusion from several unobserved systems.

The durable evidence chain is therefore longer: the owner’s preference; the entity and version to which it attached; the attachment’s retrieval; the crawler or collector’s independently verified identity; declared purpose; custody and transformation records; a model-use or exclusion decision; reviewable compliance evidence; and a remedy decision. Each link may be useful. No link inherits the authority of the previous one. That separation preserves voluntary expression while refusing to confuse an interoperable signal with the local decisions needed to make it meaningful.