Does training AI on other people’s works violate intellectual property rights?

The Trump administration prematurely dismissed the head of the US Copyright Office. Shortly before this, the Office published a report stating that using copyrighted material for training generative AI could violate the rights of authors.

Does this mean that the Trump administration wants to abolish copyright to please Musk and corporations interested in the development of AI? I would not rush to such conclusions, although, of course, as in all other cases, such drastic and insufficiently motivated cuts do not look good. But the copyright protection system does not depend solely on a particular Office director, or even on US domestic law.

Perhaps this means that there is dissatisfaction with the assessment of the use of copyrighted material for training generative AI as an infringement? Most likely, yes.

But is using copyrighted material to train generative AI an infringement? That’s where the rub is.

From the perspective of US case law, following the February decision of the Delaware District Court in Thomson Reuters v. Ross Intelligence, Americans have one case where the court of first instance has ruled that, under certain circumstances, the use of someone else’s work to train AI without the copyright holder’s permission is not a case of fair use and therefore infringes copyright. However, although this decision is authoritative, it is not yet a binding precedent and may not stand up on appeal.

Moreover, no similar practice has yet been established in other countries. For example, in EU countries, there are already court decisions where training on publicly available datasets has been recognized by the court as lawful temporary reproduction under Article 4 of EU Directive 2019/790 (Digital Single Market Directive), which establishes that text and data mining is lawful for any purpose unless the copyright holder has expressly prohibited such use of their works in advance, for example, by explicitly stating this in the terms of use of their website or product.

In Ukraine, there are currently no court decisions on a similar subject of dispute, nor is there a clear answer in the legislation regarding the legality of training AI on data sets that include protected objects of third parties. However, this legal vacuum is also a window of opportunity for the Ukrainian judiciary/legislature to form its own opinion on how to legally assess AI training on someone else’s intellectual property.

Moreover, the first signals are already being heard from the domestic doctrinal sphere that such AI training could in fact be considered a case of fair use, namely a specific form of machine reading, which is not a use of the work and therefore does not require the permission of the copyright holder. The arguments in favor of this approach do not seem so unconvincing, check it:

• From an economic point of view, forcing AI product owners to obtain licenses to use millions of works by other copyright holders is unrealistic and impossible to implement.
• If it is impossible to obtain licenses, AI owners will have to train their neural networks on works that are in the public domain (whose copyright protection period has expired). In most cases, this means works created 80-100 years ago or more. The question of the relevance of such AI knowledge is rhetorical.
• If the final AI-generated object shows no traces of direct borrowing of creative solutions from the works used, it is unlikely that the rights holders have suffered any damage.

I believe that legislative permission to freely use collections of published works for AI training is a step in the interests of the development of AI itself, rather than individual corporations, and given the importance of AI at this stage in history, a step in the interests of the overall progress of civilization.

Please do not take this as the dogmatic position of a single lawyer, but as an invitation to discussion. After all, the search for an answer to the question of legislative regulation of AI could become a touchstone for the global issue of balancing the interests of technology companies and society in the context of AI development, which is already unstoppable.