Generative AI Training and Copyright Law

Abstraction: Legal analysis of copyright liability for generative AI training data collection

Key points:

• Common practice: web-scraping copyright-protected content for generative AI training may constitute infringement
• US developers rely on "fair use" doctrine; EU developers rely on Text and Data Mining (TDM) exception
• Authors argue generative AI training fundamentally differs from traditional TDM, so TDM exception does not apply
• Training data memorization creates a separate copyright exposure pathway independent of fair use and TDM debates
• Has implications for both public academic research and corporate research on generative models
• Stober and Dornis (2025); raises the question of fair practices for communities like ISMIR (music information retrieval)

Connections: Generative AI · Training Data · Copyright Law · Text And Data Mining

Source: https://arxiv.org/abs/2502.15858?fbclid=IwY2xjawRIWGVleHRuA2FlbQIxMABicmlkETE3ZUNNcG1KeGx3aU1teGFnc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHjPtm1Wr1Q8qEgFkwTVeYwIRagiD6DgC5cSLiPRHmSD0y9Rn5Uiuk3DMIa5w_aem_FRhq8pjhQSNxK5OTjRgw-A