ONP-3100

Open News Protocol (ONP): AI Metadata

Abstract

This document defines the AI Metadata Extension: org.onp.ai-metadata, the first Extension published in this series. It covers two related but distinct concerns: disclosure of whether and how AI was involved in producing a News Object's content, and declaration of whether that content may be used to train AI models or be quoted/summarized by AI agents serving end users. It declares two Claim Domains under the mechanism ONP-3000 just established, and it is the first document in this series to be classified an Extension rather than a Companion — a genuinely different outcome from the decision test applied throughout the Companion series, not a formality.

Status of This Document

This document is part of the ONP Extension series (ONP-3000-3999). It is directly implementable. It is a Working Draft.

Normative Language

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174], per the interpretation established in ONP-0000.

1. Introduction

ONP-0000 Section 1.1 named "AI systems cannot reliably establish provenance" as one of the original problems ONP exists to address. Signature and Trust Anchor resolution (ONP-1003, ONP-0004) already solve half of that — an AI system consuming a News Object can verify who published it. This document addresses the other half: whether AI was involved in producing the content itself, and whether the publisher permits that same content to be used for training or agent-mediated consumption going forward.

2. Scope

2.1 In Scope

• org.onp.ai-metadata: fields for generation disclosure (generation_method, model_reference, human_review) and training/agent-use permission (training_permitted, agent_use_permitted, attribution_required_for_ai_use);
• the two Claim Domains this Extension declares;
• the explicit, acknowledged interaction question with the Rights Companion (ONP-2400), which ONP-3000's Claim Domain mechanism does not cover (Section 4.7).

2.2 Out of Scope

This document does NOT define:

• any technical mechanism to verify that a generation_method disclosure is accurate — like every provisional, publisher- asserted field in this series, it is trusted only as far as the publisher's own key (Section 8);
• legal interpretation of any AI regulation (e.g. the EU AI Act) or copyright text-and-data-mining exception — consistent with Principle P5, this document provides a structural signal, not a compliance determination;
• Extension-to-Companion conflict detection generally — ONP-3000's Claim Domain mechanism covers Extension-to-Extension overlap only; this document's Section 4.7 is a documented gap, not a solved one.

3. Terminology

This document introduces no new general terms beyond registering its namespace and Claim Domains (Section 4).

4. Requirements

4.1 Extension Namespace Declaration

This Extension's fields MUST be carried under onp:extensions.org.onp.ai-metadata.

4.2 Companion-vs-Extension Classification

Does "AI Metadata" have independent identity and an independent
lifecycle, separable from the Object it describes?

- AI generation disclosure and training-permission signals are
  inherently assertions ABOUT an existing Object's content. They
  cannot be meaningfully referenced or exist independently of that
  Object.
- Unlike every Companion published so far, there is no standalone
  "AI Metadata Object" a reader would ever look up on its own — it
  only makes sense attached to something else.
- NO -> Extension.

This is the first document in the series where the test's answer is genuinely "no," not a formality repeated for consistency.

4.3 Content Schema

1. generation_method, model_reference, and human_review are OPTIONAL, under the ai-generation-disclosure Claim Domain (Section 4.5).
2. training_permitted, agent_use_permitted, and attribution_required_for_ai_use are OPTIONAL, under the ai-training-permission Claim Domain (Section 4.6).

4.4 Claim Domains

Per ONP-3000 Section 4.3, this Extension declares two Claim Domains:

1. ai-generation-disclosure — assertions about how content was produced.
2. ai-training-permission — assertions about downstream AI use of the content.

Both MUST be registered in ONP-0002 alongside this document's publication (Section 10.3).

4.5 Generation Disclosure

1. generation_method, if present, MUST be one of "human", "ai-assisted", "ai-generated", or "ai-generated-human-edited".
2. Its absence MUST NOT be interpreted as "human" by default. Per the same no-default-assumption discipline ONP-2400 Section 6.2 established for Rights, an absent generation_method means undisclosed, not human-authored.
3. model_reference, if present, MUST be a string naming the AI system involved (e.g. "gpt-5", "claude-opus-4-8") — informal and publisher-asserted, not independently verified.
4. human_review, if present, MUST be a boolean: whether a human editor reviewed AI-assisted or AI-generated content before publication.

4.6 Training and Agent-Use Permission

1. training_permitted, if present, MUST be a boolean: whether this content may be used to train AI models. Its absence establishes no default assumption, consistent with Principle P5 — silence is not consent, and this document does not assert otherwise.
2. agent_use_permitted, if present, MUST be a boolean: whether an AI agent MAY quote, summarize, or act on this content when serving an end user's query — a narrower, distinct question from bulk training permission, since a publisher MAY reasonably permit one without the other.
3. attribution_required_for_ai_use, if present, MUST be a boolean: whether an AI agent using this content under agent_use_permitted MUST attribute the source.
4. This document deliberately reuses the conceptual pattern of existing text-and-data-mining reservation conventions (e.g. the TDM Reservation Protocol) rather than inventing incompatible terminology, consistent with Principle P3.

4.7 Interaction with the Rights Companion (Acknowledged Gap)

1. training_permitted and agent_use_permitted MAY appear to overlap with, or even contradict, a Rights Object's commercial_use_permitted or derivative_works_permitted flags (ONP-2400 Section 4.5) if both are present on the same Object's rights_ref and onp:extensions.
2. ONP-3000's Claim Domain mechanism (Section 4.5 of that document) detects overlap between two Extensions; it does NOT cover overlap between an Extension and a Companion, since a Companion's assertions live in content, not onp:extensions. This is an acknowledged architectural gap, not a solved one — stated here explicitly rather than left for an implementer to discover independently.
3. Pending a future resolution of that gap, an Extension asserting training_permitted or agent_use_permitted SHOULD NOT contradict what the same Object's Rights Companion states, where both are present. Where an apparent contradiction exists, a Node SHOULD apply the more restrictive reading of the two as a conservative default — favoring the rights-holder's protection over permissive interpretation when a mechanical resolution is not yet defined.

5. Object Model

{
  "onp:extensions": {
    "org.onp.ai-metadata": {
      "generation_method": "'human' | 'ai-assisted' | 'ai-generated' | 'ai-generated-human-edited', OPTIONAL",
      "model_reference": "string, OPTIONAL",
      "human_review": "boolean, OPTIONAL",
      "training_permitted": "boolean, OPTIONAL",
      "agent_use_permitted": "boolean, OPTIONAL",
      "attribution_required_for_ai_use": "boolean, OPTIONAL"
    }
  }
}

6. Processing Model

6.1 Consumption

A Node or AI system consuming a News Object SHOULD check onp:extensions.org.onp.ai-metadata.training_permitted before using that Object's content for model training, and agent_use_permitted/attribution_required_for_ai_use before quoting or summarizing it in an agent-mediated response, applying Section 4.7's conservative-reading guidance where a Rights Object is also present and appears to disagree.

6.2 Interoperability

A Node without this Extension implemented simply does not see AI disclosure or permission signals — the Object itself remains fully verifiable and usable regardless (ONP-1000 Section 4.4, rule 3). This Extension is Companion-agnostic: it MAY attach to an Article, a Media Object, or any future Companion's Object, since generation method and AI-use permission are meaningful questions for any content type, not specific to one.

7. Examples

7.1 AI-Assisted, Human-Reviewed, Partial AI-Use Permission

{
  "onp:extensions": {
    "org.onp.ai-metadata": {
      "generation_method": "ai-assisted",
      "model_reference": "claude-opus-4-8",
      "human_review": true,
      "training_permitted": false,
      "agent_use_permitted": true,
      "attribution_required_for_ai_use": true
    }
  }
}

Drafting assistance was used, a human editor reviewed the result before publication, bulk AI training is not permitted, but an AI assistant MAY summarize the article for a user as long as it cites RegioPurmerend as the source.

7.2 The Rights Interaction Scenario (Section 4.7 Worked)

Rights Object (referenced via rights_ref):
  commercial_use_permitted: false

AI Metadata Extension (same Object):
  agent_use_permitted: true
  (silent on whether the consuming AI agent's product is commercial)

Apparent tension: does agent_use_permitted=true override the
Rights Object's commercial_use_permitted=false if the AI agent in
question is part of a commercial product?

Per Section 4.7, rule 3: no mechanical resolution exists yet. A
Node SHOULD apply the more restrictive reading — here, treating
commercial agent use as NOT permitted despite agent_use_permitted's
bare `true`, since Rights' explicit commercial restriction is the
more specific and more conservative signal.

8. Security Considerations

A generation_method disclosure is publisher-asserted and not independently verifiable by Core or by this Extension — a publisher could omit or misstate AI involvement with no technical mechanism in this document to detect it, the same limitation already accepted throughout this series for any self-reported field (byline, credentials, credit). This document's contribution is giving publishers who want to disclose accurately a standard, verifiable-as- signed place to do so, not a mechanism to compel accurate disclosure from those who do not.

9. Privacy Considerations

model_reference and human_review typically carry no personal data. This document introduces no new privacy mechanism beyond what is already established elsewhere in this series.

10. References

10.1 Normative References

• [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119.
• [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174.
• ONP-0000, Introduction — Section 1.1, the AI-provenance problem this document addresses one half of.
• ONP-0001, Architecture — Section 4.4, the decision test applied in Section 4.2, here reaching Extension rather than Companion for the first time in the series.
• ONP-0003, Design Principles — Principle P3 (Ordinary Technology, Section 4.6, rule 4) and Principle P5 (Jurisdiction Neutrality, Section 2.2, Section 4.6, rule 1).
• ONP-2400, Rights — the Companion this Extension's Section 4.7 documents an acknowledged, unresolved interaction gap with.
• ONP-3000, Extension Framework — Section 4.1 (namespace registration), Section 4.3 (Claim Domain declaration), Section 4.5 (mechanical overlap detection, which Section 4.7 of this document notes does not extend to Companion interactions).

10.2 Informative References

• TDM Reservation Protocol (TDMRep) — the existing text-and-data- mining opt-out convention this document's training-permission fields are conceptually aligned with, per Principle P3.
• EU AI Act — cited as regulatory context motivating disclosure fields, not interpreted or relied upon as a compliance determination (Section 2.2).

10.3 Registry Registration

As part of this document's publication, org.onp.ai-metadata, ai-generation-disclosure, and ai-training-permission are registered in ONP-0002.

Appendix A: Full Schema Reference

{
  "onp:extensions": {
    "org.onp.ai-metadata": {
      "generation_method": "enum, OPTIONAL, no default",
      "model_reference": "string, OPTIONAL",
      "human_review": "boolean, OPTIONAL",
      "training_permitted": "boolean, OPTIONAL, no default",
      "agent_use_permitted": "boolean, OPTIONAL, no default",
      "attribution_required_for_ai_use": "boolean, OPTIONAL"
    }
  }
}

Appendix B: AI Metadata Checklist

[ ] Fields carried under onp:extensions.org.onp.ai-metadata
[ ] generation_method, if present, is one of the four recognized
    values; if absent, understood as "undisclosed," not "human"
[ ] training_permitted / agent_use_permitted, if present, are
    booleans with no assumed default when absent
[ ] if a Rights Object is also present and appears to disagree:
    apply the more restrictive reading (Section 4.7)
[ ] no claim is made that this Extension mechanically resolves
    Rights/AI-Metadata conflicts — it does not (Section 4.7)

End of Document