AI, Copyright, and the AIPPI Resolution Q295 (2025)
AI has pushed copyright into new territory, especially around the use of protected works for training and the legal status of AI outputs.
At the 2025 AIPPI World Congress in Yokohama, the International Association for the Protection of Intellectual Property (AIPPI) adopted Resolution Q295 (“AI & Copyright”), a detailed harmonization proposal on these issues.
AIPPI described Q295 as “probably the most important work that has been written on the subject to date.” (AIPPI’s own characterization: https://www.aippi.org/news/aippi-adopts-four-important-resolutions-at-the-2025-yokohama-world-congress/).
Note: Q295 is not binding law. It is a policy blueprint aimed at guiding legislators and stakeholders.
1) Training rule: authorization as the default
Q295’s baseline is clear: using copyrighted works to train AI should require prior authorization, unless an exception applies.
2) Exceptions: existing exceptions + a public-interest carve-out
Q295 emphasizes two tracks:
- General exceptions still apply (training should benefit from the same exceptions available for other uses, if legal conditions are met).
- A targeted public-interest exception is encouraged for not-for-profit training done solely for public-interest purposes (e.g., non-commercial scientific research or education). This exception should not extend to commercial exploitation of the trained system and/or the training dataset.
Where a jurisdiction allows commercial training without authorization, Q295 proposes safeguards:
- Right holders should have an opt-out, and
- If they do not opt out, a compensation mechanism should apply.
Any exception should comply with the Berne three-step test (Art. 9(2)).
3) Transparency obligations
To make rights enforceable, Q295 calls for transparency from the entity providing or training the system, including:
- Adequate information about copyrighted works used for training (so right holders can identify use and enforce rights).
- Identification of copyrighted materials input by users and used for training.
4) Outputs: how infringement should be assessed
Q295 takes several positions to clarify output disputes:
- Outputs should be assessed under ordinary infringement rules of the applicable jurisdiction.
- Style alone should not amount to infringement.
- An output should not be infringing solely because training was infringing.
- Scope matters: if authorization/exception covers only training, outputs may still infringe; if it covers outputs too, they should not.
- Moral rights remain relevant (authors may object to derogatory treatment prejudicial to honor/reputation, where applicable).
5) Liability: providers, exploiters, and “deliberate prompting” users
Depending on the facts, liability for infringing outputs may attach to:
- The provider (developer and/or entity placing the system on the market),
- The commercial exploiter, and/or
- A user who aims to generate infringing outputs (e.g., via “detailed and deliberate prompting”).
6) The AI system as an “infringing article” + remedies
Q295 proposes that an AI system can be treated as an “infringing article” where:
- More than a de minimis amount of training used copyrighted materials unlawfully; or
- The system was developed specifically to create infringing outputs.
Remedies may include (case-by-case): damages, injunctive relief, recall, and destruction, applied across:
- Infringement via training,
- Infringement via outputs, and
- The AI system itself as an infringing article.
Q295 also supports damages and/or an account of profits to address harm and the bypassing of consent, with proportionality in remedies.
Conclusion
Resolution Q295 is a granular harmonization proposal that frames AI training and outputs through familiar copyright tools: authorization, tightly bound exceptions (including the three-step test), enforceable transparency, and a tiered liability/remedy model. Even as a non-binding instrument, it is likely to influence how future rules are drafted.
For more details, please check here.