WIPO study maps five unresolved choices for AI training and copyright

The story

A new comparative study prepared for the World Intellectual Property Organization has distilled the global dispute over generative-AI training into five unresolved policy choices. Published on October 23 for WIPO’s Standing Committee on Copyright and Related Rights, the document surveys major legal approaches without recommending one. It is neither a treaty proposal nor an agreed WIPO position. Its practical value is a shared map of the questions that lawmakers, courts, rightsholders and AI developers keep answering differently.

The study, led by copyright scholar Daniel J. Gervais with contributors examining six jurisdictions, compares China, the United States, the European Union, the United Kingdom, South Korea and Brazil. It limits its scope to the use of protected works in training generative AI. Questions about whether an output infringes copyright or qualifies for protection are outside the assignment. That boundary matters because public arguments often collapse collection, training and generation into a single act even though the technology and legal analysis can differ at each stage.

Its first question is whether building a training corpus and processing it require authorization from copyright owners. Systems may gather works, normalize files, create intermediate copies and repeatedly process representations during optimization. The United States generally analyzes such uses through fact-specific fair-use doctrine, while the European Union has text-and-data-mining exceptions whose conditions and reservation mechanisms vary by purpose. The other surveyed jurisdictions supply additional combinations of exceptions, licensing rules and unresolved doctrine. The study does not declare one model compatible with every legal system.

The second choice is how a rightsholder can reserve rights at machine scale. A website notice that humans can read may not be discoverable by a crawler, while a robots.txt instruction was designed for access management rather than copyright licensing. Policymakers therefore face questions about which technical signals count, whether a reservation must identify particular uses, and whether requiring a formality-like opt-out is compatible with the Berne Convention’s ban on conditioning copyright protection on formalities. A signal can be standardized without resolving what legal consequence it carries.

Payment is the third fault line. If training is permitted subject to compensation, governments must decide what triggers payment: ingestion of a work, evidence that a model’s outputs substitute for the work, or some combination. They must also decide whether the mechanism should be voluntary collective licensing, an extended licence, a statutory remuneration right or another institution. Each choice changes who can negotiate, which uses remain possible without transaction-by-transaction clearance, and whether independent creators can realistically collect.

The fourth question is transparency. Rightsholders cannot license or challenge uses they cannot identify, but a requirement to disclose every training record can collide with trade secrets, security concerns and the scale of modern datasets. The study frames recordkeeping and disclosure as separate design problems: developers may preserve sufficiently detailed internal records while providing public summaries, regulator access or targeted evidence in litigation. The hard task is defining information that is useful enough for accountability without becoming an impractical catalogue or a route to exposing proprietary systems.

Remedies and jurisdiction form the fifth choice. A model may be trained in one country, adapted in another and offered worldwide. If a court finds an unauthorized use, removing a particular work’s influence from model parameters may be technically uncertain or disproportionate. Injunctions, damages, prospective licences and access restrictions can produce very different effects. National rules also need a theory for when domestic law reaches foreign training whose resulting service is supplied locally.

These questions will reach WIPO’s copyright committee at its forty-ninth session, scheduled for November 30 to December 4 in Geneva and online. The study’s disclaimer is important: its authors’ views are not those of the WIPO Secretariat or member states. Publication places a structured comparative record before delegates; it does not mean governments have endorsed the authors’ framing, much less agreed on a new international instrument.

INNOVOX analysis: the study’s strongest move is technical decomposition. International consensus on a single copyright exception is unlikely in the near term, but narrower interoperability may be achievable. Machine-readable reservations could use common vocabulary even when countries assign them different legal effects. Training records could share minimum fields without forcing publication of entire datasets. Such infrastructure would not settle the fair-use or compensation debate, but it could reduce avoidable uncertainty and make licensing, oversight and litigation more evidence-based.

What to watch is whether delegates convert the map into specific work. Proposals for common reservation metadata, minimum recordkeeping, or a further study of remuneration could create a path toward implementation. Member states may instead emphasize that domestic litigation and legislation must develop first. Either result would be informative. The document does not announce a global AI copyright rule; it shows why any durable rule must connect technical stages, legal permissions and remedies across borders rather than treating ‘training’ as a single undifferentiated use.

INNOVOX analysis

The document’s most useful contribution is to treat AI training as a sequence of technically distinct acts rather than one legal event. A workable international framework will need interoperable signals and disclosure practices even if countries continue to disagree about the underlying copyright exception or licensing rule.

What to watch

Watch the WIPO copyright committee’s November 30–December 4 session for member-state positions on opt-out standards, recordkeeping and remuneration. The key signal will be a concrete follow-up workstream—not language implying that a global treaty or uniform rule is imminent.