Summary Report on “Copyright and AI” (SQ 295): the Current Most Hotly Issue in IP!

This year, the Copyright Study Question addresses one of the most hotly debated issues in the field of intellectual property: the use of copyrighted works to train AI systems.

39 Reports were received in total.

All Reports can be found in AIPPI’s library at www.aippi.org.

The Summary Report can be found in AIPPI’s library here.

The following conclusions arise from the National Groups and Independent Members Reports.

A clear majority of the responding Groups consider that harmonization regarding the use of copyrighted works for training AI is desirable due to the cross-border nature of AI systems and the risks of forum shopping by AI developers.

The majority of Groups consider that prior authorization from the copyright holder should be, as a matter of principle, necessary before the copyrighted work may be used to train an AI system, unless covered by an exception. 

The majority of Groups agree that general exceptions to copyright, such as “fair use” and “temporary reproduction” exceptions, should apply to the training of AI Systems with copyrighted works, provided that the conditions to benefit from such exceptions are met.

Furthermore, the majority of Groups propose to introduce a special exception (Text and Data Mining exception) to allow the use of copyrighted works to train AI Systems without consent from the copyright holder. The majority of Groups consider that such an exception should be limited to situations of public interest, such as for purposes of scientific research, education, information or not-for-profit.

The majority of Groups also consider that exceptions should be provided for financial compensation for the copyright holder whose works have been used to train AI.

A slight majority of Groups are in favor of an opt-out system, giving the right to copyright holders to refuse the use of their works for training of AI System. In case a copyright holder doesn’t opt-out, the use of his/her works to train AI system should be lawful.

The majority of Groups believe that there should be transparency obligations of the AI developers to provide sufficient details on data/works used to train the AI system, to make it possible for the copyright holders to enforce their rights.

There is a strong consensus that an output that contains characteristic/substantial elements of a copyrighted work used to train AI, should be considered copyright infringement.  There is also a consensus that an output that is in the same style as a copyrighted work used to train AI System, should not constitute copyright infringement.   

The majority of Groups consider that the AI system itself should be considered as copyright infringement in certain specific circumstances, like, for example, when the AI system is designed to create copyright infringements, or if it systematically/massively uses copyrighted works to be trained.  

Subject to the proportionality principle, the majority of Groups consider that all traditional sanctions should be available, in case of unlawfully use of copyrighted works to train AI system, such as:

• Destruction or injunction of outputs found to be infringing.
• Destruction or injunction of the AI system itself, but only under specific conditions, including (i) the circumstances in which the AI system was trained (e.g., bad faith/awareness of the infringement, sufficient diligence taken to avoid infringement, AI developed to infringe copyright);and (ii) the gravity of the infringement (e.g., amount/proportion/value of the infringing works used).
• Destruction of copyrighted works that have been unlawfully used to train an AI system and the prohibition of its further use for training purposes. In case such destruction is not possible, alternative measures like financial compensation should be imposed.

Next step: preparation of the First Draft Resolution by the Study Committee Leadership. See you in Yokohama!