Training GAI models involves the extraction, copying, and analysis ‒ in an opaque manner (blackbox) ‒ of vast amounts of data, often protected by copyright. Artists and authors are concerned about their works being used without authorisation, remuneration, or credit to create new artificial works.
Adopted on 21 May 2024, the European Artificial Intelligence Regulation (AI Act) aims to address the many risks posed by AI while fostering innovation. It will enter into force in August 2024 and will be fully applicable by summer 2026. The ban on certain AIs will already be applicable by February 2025, and the obligations for gpAIs by Summer 2025. Companies would be wise to start preparing now.
What’s new in terms of copyright?
-
1. Data training is authorised for scientific research, subject to opt-out in other cases
The text and data mining (TDM) exception provided by Directive 2019/790 is fully applicable to GAIs. TDM is authorised for scientific research purposes but is subject to the absence of reservation of rights (opt-out) from copyright holders for other purposes.
-
2. Copyright policy is mandatory for general-purpose AI providers
Providers of general-purpose AI models (gpAIs, including GAI) must implement a policy to comply with copyright law, particularly to identify and respect opt-outs, including through state-of-the-art technologies.
-
3. Summary about the content used for training is also mandatory
GpAIS must make publicly available a sufficiently detailed summary about the content used to train the models, according to a simple and effective template provided by the European AI Office (freshly created within the European Commission).
The summary should be comprehensive in scope rather than technically detailed to facilitate copyright holders in exercising and enforcing their rights, e.g. listing the main data collections or sets (e.g. databases or data archives) and providing a narrative explanation about other data sources used.
-
4. Exceptions yet apply for nonprofessional, scientific uses and SMEs
Individuals who develop or use models for nonprofessional or scientific research purposes are not obliged but nevertheless encouraged to voluntarily comply with these requirements.
SMEs, including start-ups are given simplified ways of compliance that should not represent an excessive cost and not discourage the use of gpAI models. gpAIs placed on the market after August 2023 are given until August 2027 to comply with the Act.
-
5. The Act applies whether established in the EU or elsewhere in the world
Providers that place gpAIs on the EU market should ensure compliance, regardless of where the data training takes place and whether those providers are established or located within the EU or elsewhere.
-
Conclusion
Compliance with these obligations will be monitored by the AI Office, although it will not conduct an assessment on a work-by-work basis. It will be up to copyright holders to verify compliance and take action for the rest.
Unless made accessible under a free and open-source licence, providers of gpAIs must also draw up (i) a technical documentation about their training and testing process and evaluation results; and (ii) information and documentation to supply to providers that intend to integrate the gpAIs model into their own AI system.
If the free and open licenced gpAI model poses systemic risks, other obligations on risks, incidents and cybersecurity apply, compliance with which depends on adherence to codes of practice. Codes of practice must be prepared by May 2025.