Copyright infringement in generative AI
Following our exploration of copyright protection for AI-generated content, we now turn to another big issue in the digital age: the potential liability for copyright infringement by generative AI. As discussed in the previous analysis, even if the output of GenAI models qualifies for copyright protection, this is not a guarantee that it does not also violate third parties’ rights — including copyrights, trademarks, or personality rights — if produced without proper authorization.
This is because GenAI models derive their output from extensive datasets, which often include copyrighted works. Although these models are capable of producing a vast array of new materials, their outputs sometimes closely resemble existing works or draw heavily on cultural references. Yet, the legal implications remain unclear: when is the output “inspired” from previous works, and when is it actually “copying” them? Where to draw the limit?
In July 2023, a notable legal confrontation unfolded when a class-action lawsuit was brought against Meta. Plaintiffs alleged that the output from Meta’s large language model, LLaMa, constituted infringing derivative works. The court, however, dismissed these claims as “nonsensical” due to the lack of evidence that LLaMa’s outputs were adaptations of the plaintiffs’ literary works. This dismissal underscores the complexities of proving copyright infringement in the context of AI-generated content.
At the same time, as GenAI models become adept at mimicking artistic styles and techniques of established artists and designers, the likelihood increases that their outputs could be seen as infringing on the copyrights of these third parties. If substantial similarity is established, anyone using the AI-generated content could be held liable for copyright breach. The strict liability regime of copyright infringement, which holds individuals accountable regardless of intent, raises concerns about the fairness of such legal actions.
Vicarious liability becomes particularly relevant when GenAI models generate works that infringe on copyright without the user’s knowledge — or without means of knowing — about the potential violation. This form of liability hinges on the principle that if an entity benefits financially from an activity that infringes copyright and has the ability to control or prevent that activity, it should take adequate steps to mitigate infringement risks. Therefore, AI providers, by offering tools capable of generating potentially infringing content, could face legal consequences even without direct involvement in the creation of such content. An illustrative case is the ongoing lawsuit against Stable Diffusion, where AI firms are claimed to be vicariously liable for copyright violations.
Even if style itself is not protected by copyright, using GenAI models to create works in the style of existing artists or creators without their explicit approval could implicate moral rights. Under EU copyright law, the moral right of integrity allows authors to prevent the use or dissemination of works associated with them — if these works compromise their artistic vision or personal beliefs. Therefore, AI-generated content that incorporates elements or themes disapproved by the original artists could potentially infringe on their moral right of integrity.
Beyond copyright, questions arise regarding the adequacy of copyright protection to shield against unfair competition and commercial practices. GenAI’s ability to create images, music, or videos impersonating public figures or mimicking the voices of renowned singers without consent brings up significant issues concerning personality rights and authenticity. For example, the case of the song ‘Heart on My Sleeve’ involved AI imitating the voices of Drake and The Weeknd without permission, leading to a request for its removal from major platforms by Universal Music Group. Although the underlying music was new, the recognizable use of these artists’ voices raised complex issues in copyright and personality rights.
In some instances, personality rights — which protect against unauthorized use of an artist’s voice or likeness — might provide a legal basis to protect artists’ rights when conventional copyright law falls short. Additionally, the misuse of personal data, including artists’ voices, could lead to data protection infringements under regulations like the GDPR or be considered a criminal offense under the European member states.
As we navigate this evolving landscape, it becomes clear that unfair competition law might also play a crucial role. Leveraging an artist’s distinctive style without authorization to compete in the same market could be viewed as unjust exploitation of their creative efforts. These practices, while not directly infringing copyright, challenge the fairness of competitive practices in the creative industries.
The upcoming articles will continue to delve deeper into the intricacies of copyright training data datasets, and the legal status of GenAI models, proposing possible solutions to address these legal complexities.
