WIPO Study on AI Training and Copyright: Five Open Questions
Before WIPO's Standing Committee on Copyright and Related Rights (SCCR) meets for its 49th session, one document in the meeting papers deserves a close read. On 23 October WIPO published the Scoping Study on AI Training and Copyright (SCCR/49/6) as a document for SCCR/49, to be held in Geneva from 30 November to 4 December 2026. The study was commissioned by the WIPO Secretariat, and its disclaimer says the views are the authors', not those of member states or the Secretariat.
It answers a request made at SCCR/47 (1–5 December 2025) for a study on policy or regulatory approaches to AI training and copyright, and on practices for authorisation, enforcement and compensation. The lead co-author is Daniel J. Gervais, emeritus professor at Vanderbilt Law School; the other authors come from the Republic of Korea, Kenya, Brazil and China, including Wang Qian of East China University of Political Science and Law.
The study is deliberately restrained. It describes and compares approaches in jurisdictions including China, the US, the EU, the UK, the Republic of Korea and Brazil, without recommending any of them or taking sides on disputed issues, and it covers only the training of generative AI; whether outputs infringe or can be protected is outside its brief. Its three parts cover how training uses works (data gathering, training, whether copies remain in the model); authorisation, licensing and compensation (individual and collective licensing, text and data mining exceptions, remuneration); and enforcement (whether a model is a copy, jurisdiction over training abroad, transparency and opt-out standards).
The conclusion offers no answers but sets out five choices member states face:
- whether reproductions made to build a corpus and train a model need authorisation or fall under an exception or limitation (AI-specific, a TDM provision or a general clause), and on what conditions, such as lawful access, purpose and commercial use;
- how any reservation (opt-out) mechanism should work, given that machine-readable reservations may be beyond individual authors while reservations that cannot be acted on at scale leave developers no compliance path, and whether it is compatible with Article 5(2) of the Berne Convention;
- whether compensation attaches to ingestion, to output substitution or to both, whether it takes the form of licensing built on exclusive rights or a statutory remuneration claim, and through which institutions it flows;
- what record-keeping and disclosure of training material is needed to exercise rights and administer remuneration, reconciled with trade secrets;
- whether remedies stay effective and proportionate when works cannot be removed from a trained model, and how jurisdiction works when training happens abroad but services are offered domestically.
J&C notes that this is a study, not a draft treaty, and it changes no country's current rules; its value is that it frames the coming SCCR discussion around these five questions. Licensors and users of training data can check their existing licences against the five points, looking at how they deal with training use, opt-outs, record-keeping and disclosure, and compensation, and follow the discussion at this SCCR session.
Source: www.wipo.int | www.wipo.int
