AI Design and Copyright

What would design by artificial intelligence mean for the law and practicalities of copyright?

In "Judging a Book by its Cover", Brian Kenji Iwana and Seiichi Uchida of Kyushu University report an elegant study into the use of a "Convolutional Neural Network (CCN) to predict the genre of a book based on the visual clues provided by its cover". Having learned from over 100,000 book covers from Amazon.com in 20 categories, the neural network was able to place new book covers within the correct category 20% of the time. Based on these initial results, the authors "hope to design a network to better capture the essence of cover design."

MIT's Technology Review's write up of the paper concludes "This is interesting work that could help designers improve their skills when it comes to book covers. A more likely outcome, however, is that it could be used to train machines to design book covers without the need for human input. And that means book cover design is just another job that is set to be consigned to the history books."

"book cover design is just another job that is set to be consigned to the history books" –MIT Technology Review

There are already examples of this trend in digital books, such as the pseudo-leather bound covers frequently used for eBooks of out-of-copyright classics or these more elegant solutions by Mauricio Giraldo Arteaga of the New York Public Library Labs:

Without comment on the disconcerting social, cultural and economic impacts of such AI taking over design, here are a few thoughts on the implications for copyright (based on my experience as a UK-based Intellectual Property lawyer).

• It is already possible to own the copyright in works generated by computer, and this should include book covers (i.e. artistic works). The "author", and hence the default owner of any copyright, is "the person by whom arrangements necessary for creation of the work are undertaken".
• It is more questionable whether computer generated book covers would satisfy the requirement that artistic works be "original", i.e. the product of the author's independent skill, labour, taste, judgment and "own intellectual creation". It has been suggested that a computer-generated work should be assessed as if it had been done by a human author. Such an approach would be likely, for example, to protect covers of the standard of Mauricio Giraldo Arteaga's above. Time will tell whether covers designed by AI would satisfy such a test and whether such a test would be attractive to the Court or future legislators on public policy grounds.

It is unclear whether AI cover designs would be copyright infringements at all or infringements worth pursuing.

• An automated process can produce liability for copyright infringement (hence, for example, the many claims against Google for copyright infringement based on the outcomes of its automated "bots" and search results).
• A cover design generated by a neural network will be derived from the covers that the network has analysed. However, this may not amount to infringement of copyright in the covers analysed and, in reality, creative and "original" human designers inevitably derive their own creations, in part, from their experience of other designers' work.
• First, if the neural network derives rules of design, rather than copying specific parts of actual covers, it may avoid infringement on the general principle that copyright protects the expression of ideas and not ideas (or rules) themselves.

• Second, infringement requires copying of a substantial part of a copyright work. Traditionally, this applies whether or not the part copied becomes a substantial part of the infringing work. On that basis, a system that – for the sake of argument – literally averaged thousands of copyright works would infringe all the underlying works. However, some cases have expressly required a visual similarity between the part copied and its appearance in the resulting work. The intriguing works of Idris Kahn are often created by combining multiple images, such as every page of the Quran or every late Constable painting or every stave of Chopin's Nocturnes. I include an example above (Idris Khan, every ... Bernd and Hilla Becher Prison type Gasholders, 2004), which layers a series of images including those below (Bernd and Hilla Becher, Gas Tanks, 1965–2009). Is Khan's image visually similar to any of Bernd and Hilla Becher's? The general shape is the same, but the visual impact of the layered image is radically different from any one of the source images. Any similarity, questionable as it is, results from the unusual uniformity of the source images. For cover designs based on thousands of sources, even this limited similarity would not occur. However, A neural network capable of convincing book designs may need to cherry pick parts of actual covers, rather than averaging multiple works, and there may be real risks of copyright infringement.

• In any case, an AI system may tend to pick relatively generic aspects of source works and this would limit the practical likelihood that any copying would be detected or be considered worth enforcing or merit substantial compensation. Like human designers, AI systems should avoid copying distinctive elements of individual copyright works. Theoretically, an AI system could be just as capable (or even more capable) of identifying unusual design elements and avoiding high risk infringement.

If AI design does become commonplace, some practical measures might be considered to reduce its impact on Intellectual Property:

• Just as websites can prevent search engines indexing them or copying their contents via options in their robots.txt (explained here), owners of copyright works might seek to use equivalent technical measures to prevent neural networks feeding on their copyright works.
• Significant sources of copyright works, such as Amazon, might be encouraged to watermark covers and other copyright images (though publishers may prefer to run the risk of infringement to avoid marring the appearance of their wares).
• There is a risk of AI being used to generate large amounts of IP in an attempt to create abusive revenue-generating portfolios, particularly for "monopoly" rights which can be infringed without copying. This risk is already minimised for some rights by examination of validity and high registration costs. These limit the scope to exploit the patent system but other rights, especially registered designs, might need reforms to avoid a flood of AI-generated registrations.
• AI may itself be employed to search for infringements (by human or artificial designers) drawing on large databases of copyright works. There are already programs to search for exact copies of copyright works (e.g. the aptly named FireFox add-on "Who stole my pictures?"); AI could look for subtler forms of prima facie copying.

Matt Hervey is a "tech-obsessed" Intellectual Property specialist at Gowling WLG, based in London.