A new court decision in the US is making the AI industry nervous – several lawsuits are pending against developers and operators of (generative) AI services. They are accused of using copyrighted works without permission to train their systems and, on the other hand, their output is sometimes so similar to the training material that it creates competition causing economic harm to the copyright holders.
In their defense, AI companies often invoke the American Fair Use doctrine, which states that in some cases it is permissible to use copyrighted material without permission, for example, in the context of reporting, science, etc. For example, in the large-scale lawsuit by the American music industry against the AI music generators Suno and Udio, the defendants refer to Fair Use, among other things.
The extent to which unauthorized use is tolerated under Fair Use (in America; in Germany we have no equivalent) depends on various factors and is decided on a case-by-case basis. Important considerations are whether the use is commercial or not, whether it is transformative, and whether it impairs the value of the protected works.
In a copyright case that has been ongoing since 2020, a judge in Delaware has now ruled, contrary to his original opinion, that the use of copyrighted material (short summaries of legal texts from the Westlaw platform) as a basis for an AI-based search engine did not fall under Fair Use in the case under review, since ultimately a competing product was developed that potentially causes economic harm to the original copyright holders. The plaintiff was copyright holder Thomson Reuters against AI startup Ross Intelligence (for more details on the case see Heise, or court decision).
Although in this case, no generative AI was yet used in the unauthorized use; as far as we can understand, the AI algorithms working in the background are actually secondary to the core problem. Nevertheless, the big question now is whether other courts will follow the negative argumentation regarding Fair Use, provided that a copyright infringement is previously established in the proceedings against the relevant AI services.
It is an open secret that most large, generative AI models are trained with little regard for the correct licensing of the works used. Recently, in another court case, it came to light that Meta trained its Large-Language Model Llama, among other things, with the pirated LibGen dataset, and with Zuckerberg&s explicit approval.
