[Theme] Digital Rights Infrastructure for Text: A Production Use Case for Agent Identity

Theme Name

Digital Rights Infrastructure for Text: A Production Use Case for Agent Identity

Proposer(s)

Alexandre Leforestier (Panodyssey)

Summary

Panodyssey is a European company in the publishing and media sector, subject to EU digital law. It operates AI governance and access rules, from the highest level (domain) down to the lowest level (content ID), today made up of five layers in production, compliant with applicable European law (Copyright DSM Directive, AI Act):

1. AI governance and domain-level access rules (TDM-REP, Content-Usage);
2. a machine-readable discovery layer, per publication;
3. a structured rights declaration per work (ODRL/JSON-LD), distinguishing indexing, RAG and training;
4. an auditable, timestamped registry of every rights declaration, with a full evidentiary chain;
5. certified author identity: physical and legal persons authenticated per publication (dual KYC).

This chain works end to end on the publisher's side. But it stops at the agent's door: the agent side remains today totally anonymous from a legal standpoint. Even when an agent can be recognized by its self-declared user-agent string, that label has no legal value: it is declarative, unverifiable by a third party, and does not constitute an identity that can be held accountable at the point where the contract must be enforced (compliance or non-compliance with the intellectual property rights declaration: a certified IP registry, publication by publication, AI by AI, author by author, content format by content format).

This is the problem to solve: the entire rights chain can be correct and yet remain without effect, for lack of an agent identity to hold it against. This observation concerns text here; it is duplicable to other content formats, whether related or not (image, audio, video).

Scope / Open Questions

• Use case: an AI agent accesses text content, published online or offline but accessible via API or dataspace. At the point of access, and with no prior relationship with the publisher, it must be able to determine whether the intended use is authorized, and under what conditions. Conversely, the author, publisher, or their representative must know who is using this content and be able to track it legally and economically.
• Architectures: an auditable, timestamped registry of rights declarations, linking each declaration to a publication, an AI, an author and a content format, is a reference architecture already deployed on the publisher's side. The symmetrical building block is missing on the agent's side: a verifiable identity.
• Trust framework: a contract requires legal representation of both parties. On the publisher's side, this representation exists (individual/corporate KYC). On the agent's side, it does not yet exist, independently of how the agent's own legal personhood will eventually be settled.
• Machine-readable trust metadata: the audit registry described above (timestamp, previous value, new value, content identifier) is an operational example of this kind of metadata, on the rights-declaration side; it would be worth comparing with the group's work on agent identity.

This work also aims to encourage transparency on all the components of a piece of content, as well as on its context of creation, and to give value to ethics and human creation. This is in line with the cross-cutting discussion at the 1st meeting on ethics and human rights, in particular the "human-rights-by-design" approach.

Proposed contributions:

1. Share the audit registry as a real use case (a production corpus, not a synthetic example) to test the group's proposals on agent identity and attestation.
2. Propose a technical liaison note with the IETF (aipref), to bridge their work and the group's work on agent identity.
3. Contribute to a taxonomy of evidence/rejection classes (valid and current declaration, replayed, expired, malformed, correctly signed but out of mandate) based on the operational experience described above.
4. Document the correspondence between this chain's five layers and the group's trust framework, as a sector-specific case study for text. In this reading, layer 6 would be agent identity, the very subject of the group's work, the missing piece that makes the first five layers enforceable. Layer 7 would ensure interoperability with IETF content-preference standards (aipref), so that the rights declaration remains readable and consistent beyond Panodyssey's own scope alone.
5. Facilitate dialogue between AI systems, creators or their representatives, and other stakeholders, toward a concrete collaborative goal.

Related Work

• Directive (EU) 2019/790 (Copyright DSM) and Regulation (EU) 2024/1689 (AI Act)
• IETF draft-ietf-aipref-vocab (content preference vocabulary: indexing, training, summarization, reproduction; does not cover agent identity or rights enforcement/verification)
• TDM-REP / TDM-Policy (W3C)
• Creative Commons (established complementarity: CC licenses = human-readable signal, AI Notice = machine-readable signal, independent layers)
• UNESCO Media and Information Literacy (MIL) Alliance
• EUIPO (European observatory on infringements of intellectual property rights)
• Ongoing European consortia and projects on content and AI governance (including the CREA Trust AI consortium, Creative Europe)

Related Themes

This theme complements, on the content side, the group's work on agent identity and attestation. The publishing and media sector illustrates this need well: its content, its uses and its rights chain have specific characteristics that call for a sector-specific response, complementary to the group's more general approaches.

Discussion