Ryan Ward: Is copyright law well-suited for advances in artificial intelligence?

A book was recently published on lithium-ion batteries. You may wonder why this is the topic of a tech law article, especially with one of the simpler sentences in the book being:

“Due to its high theoretical capacity (718 mAh g−1 ), low cost, relative abundance, and environmental benignity, NiO has attracted considerable attention among multiple TMOs for Li-ion batteries.

The reason is that, apart from the preface, this 278-page book was written entirely by a machine learning (ML) program – a virtual author named Beta Writer. While this is interesting in many ways, it has significant implications from an Australian copyright law perspective[1].  In particular, who owns the copyright in a book generated by a ML program and does copyright even exist in this book?

While there are a number of requirements that must be met for copyright to subsist, the one that is likely to be tested by the advent of ML is human authorship. Copyright law requires that the work be the result of some exertion of intellectual effort by a human author. If the work actually comes from a computer or program, and the involvement of a human is only minor, and is not essential to the development of the work, copyright won’t subsist in the work.

For the book above, it is possible the developers of the ML program would not have had sufficient input into the final product that was produced to be considered authors. The program mined SpringerLink, a database for journals and textbooks, and condensed thousands of papers into a book. Apart from pointing the program to this database and using state of the art natural language processing, the authors were not involved in the actual curation or summation to create the book. This means that the creators of the ML program will have an uphill battle in arguing that copyright subsists in the book.

At its core, copyright protection is about protecting the expression of original works from unapproved derivation, giving the owner sufficient economic incentive and reward to continue to create. In contrast to tangible property, copying intellectual property can be extremely cheap compared to the cost of producing the original. Without IP protections like copyright, copiers could undercut producers, disincentivising creation altogether.

This is concerning as it may limit the production of ML work that provides a benefit to society. For example, ML programs are capable of compiling and summarising research that extends over many years and across disciplines. This allows for a low cost method of circulating ideas and knowledge amongst not only academics and practitioners, but potentially also to laypeople who have an interest in the area.

What does this mean for this ML-generated publication? It depends on if you believe that publication can be incentivised at scale without sufficient IP protections. While these authors may be incentivised by a culture of open access, this is likely to be only a fraction of the community and limits the contributions of authors who rely on the creation and publication of copyright works for their income. Perhaps contract or trade secrets could go some of the way, but if we want to extract the maximum value from artificial intelligence, it is likely that copyright law needs to accommodate all forms of authorship, even if that author could be part of the artificial intelligence revolution one day.

[1] This book is published in the USA, so Australian copyright law does not apply. This article focuses on the limitations if it were to be published in Australia.