Copyright in AI model training
How it works
The analysis runs along several tracks depending on jurisdiction. (1) In the US a court weighs the fair-use defence under four factors โ the purpose and character of the use (including transformativeness), the nature of the work, the amount used, and the effect on the work's market; often decisive are whether the model produces outputs that compete with the original and whether verbatim fragments are reproduced (memorization). (2) In the EU the Article 4 TDM exception of Directive 2019/790 applies by default unless the rightholder has effectively reserved rights (opt-out in an appropriate, e.g. machine-readable, manner); Article 3 grants a broader exception to research organisations. (3) The AI Act requires GPAI providers to maintain a copyright-compliance policy and publish a summary of training data, strengthening enforceability of the opt-out. (4) An alternative to relying on exceptions is data licensing (deals with publishers and content databases). Disputes lacking a clear defence go to court and end in partial rulings or settlements.
Problem solved
Addresses the legal uncertainty around training data: whether mass copying of works to train models requires consent and licensing or falls within exceptions (fair use / TDM), how to allocate liability for infringement, and how creators can control the use of their works.
Components
US doctrine assessing the legality of unauthorised use of a work under four factors (purpose and character, including transformativeness; nature of the work; amount; market effect). Central to US training disputes.
Exceptions in Articles 3 and 4 of Directive 2019/790: Article 3 for research organisations and cultural-heritage institutions, Article 4 a general exception permitting mining unless the rightholder has reserved rights.
Acquiring rights to training data through agreements with publishers, agencies and content databases โ a route that limits litigation risk instead of relying on fair use/TDM.
The question of who is liable for infringement โ at the training stage (reproduction) and at model output (reproducing works, memorization): the model developer, the data provider, or the user.
The rightholder's ability to reserve, in an appropriate (e.g. machine-readable) manner, that their works may not be mined, disabling the general Article 4 TDM exception and reinforced by AI Act obligations.
Evolution
The EU adopts the Directive on Copyright in the Digital Single Market, introducing text-and-data-mining exceptions, including a general exception with an opt-out option.
In January/February 2023 artists (Andersen et al.) and Getty Images sue Stability AI over the use of works and photos to train Stable Diffusion.
In December 2023 the NYT files suit, arguing that model training and outputs infringe copyright and compete with the publisher's business.
In June 2024 the RIAA and major labels sue the music generators Suno and Udio for training on recordings without consent.
The AI Act enters into force in 2024; it obliges general-purpose AI model providers to maintain a copyright-compliance policy (including the opt-out) and publish a summary of training data.
In February 2025 Judge Bibas rules that Westlaw headnotes are protectable and Ross's use of them for training (a non-generative model) is not fair use โ the first significant ruling on the merits.
In June 2025 Judge Alsup finds training on lawfully purchased books to be fair use, but copying pirated copies is not; Anthropic reaches a settlement of about $1.5 billion.