22  Appendix 2: Preference Signals for AI Data Use

Status: first draft complete.

22.1 Overview

This appendix provides a brief description of, a links to, information on emerging “AI Preference Signaling” from Creative Commons and the IETF (other initatives and orgs may be added as well).

Key links:

• “CC Signals: A New Social Contract for the Age of AI”
• “CC Signals Implementation”
• “creativecommons/cc-signals”
• https://www.ietf.org/archive/id/draft-ietf-aipref-vocab-02.html

What CC signals are: A Creative Commons framework for reciprocal AI reuse: content stewards can allow specific machine uses if certain conditions are met (e.g., credit, contributions, openness). Overview & implementation notes.

• Four proposed CC signals (v0.1)

  • Credit (cc-cr) — cite the dataset/collection; RAG-style outputs should link back when feasible.
  • Credit + Direct Contribution (cc-cr-dc) — proportional financial/in-kind support.
  • Credit + Ecosystem Contribution (cc-cr-ec) — contribute to broader commons.
  • Credit + Open (cc-cr-op) — release model/code/data to keep the chain open. Source (draft repo & posts).

• IETF AI Preferences (aipref) — the transport & vocabulary

  • Vocabulary: a machine-readable set of categories (e.g., ai-use, train-genai) and preferences (y = grant, n = deny) with exceptions. Drafts.
  • Attachment: how to convey these preferences via HTTP Content-Usage header and robots.txt extensions. Drafts.
  • Structured Fields: uses RFC-standardized HTTP structured field values.
  • Robots Exclusion Protocol baseline.

• Putting them together (content-usage expression)

  • Shape:

    <category>=<y|n>;exceptions=<cc-signal>

    Example in robots.txt (allow everything, but AI use denied unless Credit):

    User-Agent: *
    Content-Usage: ai-use=n;exceptions=cc-cr
    Allow: /

    Example HTTP header (deny gen-AI training unless Credit + Ecosystem):

    Content-Usage: train-genai=n;exceptions=cc-cr-ec

    (Syntax and examples from CC & IETF drafts.)

• Operational notes (for this repo’s flywheel)

  • Per-record fields to store: license (CC0/CC-BY/CC-BY-SA) and ai_pref (IETF aipref value + optional CC signal), plus optional attribution handle. (Aligns with CC write-ups & IETF drafts.)

  • Placement:

    • Location-based signals via robots.txt for site/paths.
    • Unit-based signals via HTTP Content-Usage on dataset files and API responses.

  • Interoperability expectations: signals are normative preferences; adherence relies on ecosystem norms (similar to robots.txt & CC license culture).