Artificial Intelligence and Intellectual Property: Copyright, Idea-Expression Dichotomy, Merger Doctrine and Fair Use
An argument for the use of the works of copyright holders by artificial intelligence is the idea-expression dichotomy. The idea-expression dichotomy separates an idea from expression of the idea. For this reason, it is often argued by proponents of the use of copyright holders works in artificial intelligence training models that the models do not reproduce or redistribute the expressive content of training data. Instead, like search engines or text-mining tools which analyse copyrighted materials without infringement, they distill non-expressive insights. However, a lot of training for artificial intelligence models involves copying. The idea-expression dichotomy doesn’t automatically immunise the intermediate copying step. Issues arise where there is exploitation of protected expression.
The merger doctrine makes an expression uncopyrightable when an idea can only be expressed in a very limited or single way. In view of this doctrine, questions could be raised as to whether artificial intelligence outputs avoid infringement by not copying merged expressions.
US cases (such as Bartz v. Anthropic, Kadrey v. Meta) have found training “exceedingly transformative” and fair use when data is obtained lawfully. In other cases such as Thomson Reuters v. Ross Intelligence, fair use was rejected in competitive contexts. Fair use outcomes can vary. In 2025, the US Copyright Office released a report on the copyrightability of works generated by artificial intelligence. The report concludes that there is no single answer as to whether fair use can be found where there is unauthosed use of copyright materials to train AI models.
The contrasting approaches of the US with other countries like the UK is revealed where fair use is applied. In the UK, fair dealing is an exception to copyright law, allowing for the use, without licensing, of copyrighted works in particular circumstances. Without a commercial text and data mining exception or license, there is risk of infringement when training artificial intelligence using reproduction of original works.
Disclaimer: This publication is for educational and informational purposes only and does not constitute formal legal advice or create an attorney-client relationship