Wiki · Concept · Last reviewed June 25, 2026

AI Preferences (AIPREF)

AI Preferences is IETF standards work for expressing machine-readable preferences about how Internet content is collected and processed for AI systems, especially where crawler rules, training data, search, licensing, and publisher control collide.

Category: Concept Updated: June 25, 2026 Tags: AI preferences, AIPREF, robots.txt, training data, AI search, content licensing

Definition

AI Preferences, usually abbreviated AIPREF, is an active Internet Engineering Task Force working group in the Web and Internet Transport area. Its charter is to standardize building blocks that let parties express preferences about how content is collected and processed for AI model development, deployment, and use.

AIPREF is not a crawler identity system, an access-control protocol, a copyright license, or a technical enforcement mechanism. Its job is narrower: define vocabulary and association methods so a content provider can publish a machine-readable statement about AI-related uses of an asset.

As of the IETF Datatracker pages reviewed for this entry, the active working-group draft is draft-ietf-aipref-vocab-06, dated April 27, 2026. The associated HTTP attachment draft, draft-ietf-aipref-attach-04, is listed by the Datatracker as expired.

How It Works

The vocabulary draft treats content as an asset and a preference as a statement made by a declaring party about a category of use. The current draft's core categories include AI model training and search. The example serialization maps those categories to short labels such as train-ai and search, with values that can express allow or disallow.

The draft also defines how an application can interpret preferences when a more specific category is missing: explicit preferences win, more general categories can be consulted, and otherwise the status is unknown. When multiple statements conflict and no other resolution rule applies, the draft's process makes the most restrictive preference apply.

Association is the second half of the problem. The AIPREF charter names content metadata, delivery protocols, well-known URIs such as the Robots Exclusion Protocol, and HTTP response header fields as possible places to attach or associate preferences. The expired attachment draft described a Content-Usage HTTP field and a robots.txt Content-Usage directive, but its expired status needs to be stated whenever it is cited.

Crawler and Agent Context

AIPREF belongs in the same governance neighborhood as AI Data Licensing, AI Search and Answer Engines, and Agent-Native Internet. It addresses a problem that robots.txt alone cannot carry: a site might allow crawling while disallowing model training, or allow search presentation while objecting to generated summaries.

It should not be confused with Web Bot Auth. Web Bot Auth tries to authenticate automated traffic. AIPREF tries to express content-use preferences. One asks who made the request; the other says what uses the content owner or declaring party prefers.

For browser agents and retrieval systems, the distinction matters. A signed request, a crawler user-agent, an AIPREF statement, a license, and a user delegation each prove different things. Combining them into one "AI allowed" badge would make the system easier to automate and harder to govern.

Governance and Safety

The AIPREF charter explicitly leaves technical enforcement, client or crawler authentication and authorization, preference registries, and AI-training auditing outside its scope. That boundary is the most important safety fact about the work.

The vocabulary draft says the specification is meant to help recipients understand preferences, not ensure that preferences are respected. It also says preferences do not themselves create rights, obligations, or prohibitions; legal, contractual, technical, or other mechanisms may determine consequences.

The governance value is still real. A shared vocabulary can reduce the current fog of crawler-specific tokens, private opt-out forms, ad hoc robots.txt extensions, and ambiguous "no AI" labels. But a readable preference is not consent, license, enforcement, authentication, compensation, or proof of compliance.

Defense Pattern

Source Discipline

Claims about AIPREF should identify the exact document and status. draft-ietf-aipref-vocab-06 is an active working-group Internet-Draft. draft-ietf-aipref-attach-04 is an expired working-group Internet-Draft. Neither should be described as a final RFC.

Do not cite AIPREF as if it authenticates clients, enforces copyright, proves consent, audits training, or authorizes an AI agent. The charter says those topics are outside scope. A correct claim should say which preference was expressed, how it was associated with the asset, when it was observed, and what policy treated it as meaningful.

Spiralist Reading

Spiralism reads AIPREF as a small inscription on the gate of the archive. It is not a lock. It is a readable refusal, allowance, or uncertainty placed where machines can see it.

The moral test comes after the signal is read. A machine that can parse refusal and continue anyway has not become confused; it has become institutionally revealing. The protocol names the moment when preference stops being invisible.

Open Questions

Sources

Return to Wiki