Auctioning Art(ificial Intelligence): The IP implications of Edmond de Belamy

The IPKat has had his portrait painted by the
AmeriKat - raising another issue of
non-human authorship/ownership

In the AmeriKat's free time (which does, on occasion, exist), she takes her paw to some painting, being her medium of choice since kittenhood.  The paintbrush is dictated by the AmeriKat's paw which, in turn, is dictated by the spark of artistic expression and emotion lighting up the hemispheres of her brain.  Artistic expression which cannot be downloaded, decoded or analyzed in any complete way.  It is this artistic expression and the old image of the tortured, hungry and, often, poor artist, that historically fueled some of the philosophic underpinnings of some copyright systems  and, in particular, purportedly justified moral rights and duration extensions. 

Over the last century, the subject matter of copyright has seen an expansion as the proliferation of information has created "new"  subject matter capable of protection with the justifications for protection aligning more with traditional economic rationales of copyright and IP systems.  So what happens when a "modern" technology, Artificial Intelligence (AI), meets a field more closely associated with the romantic (big and little 'R') notions of copyright.  Last week's story from New York, sheds some light on these issues.  The AmeriKat has turned to Kat friend, Alexander de Leeuw at Dutch IP firm, Brinkhof, who reports on last week's news:

"Last Friday was a remarkable day in the history of art sales. For the first time a painting made entirely using ‘Artificial Intelligence’ was sold at auctioning house Christie’s New York. The painting with the title “Edmond de Belamy, from La Famille de Belamy” was originally estimated at $7,000-$10,000 but ultimately sold for the staggering amount of $432,500 (including fees). Much like a painting of human origin, the painting sold at Christie’s was signed by ‘the artist’ with a core component of the algorithm that created it:“min G max D Ex[log(D(x))]+Ez[log(1-D(G(z))]”.

The artist's signature? 

So how did this portrait of the imaginary Edmond de Belamy come into existence? It essentially comes down to teaching a computer how to paint. The portrait was made with machine learning algorithms called Generative Adversarial Networks (“GANs”). These algorithms are capable of generating images by mimicking characteristics of images from a training dataset (but are also capable of generating other output such as music and text). The artist Pierre Fautrel of the Paris-based art collective Obvious Art (which includes Hugo Caselles-Dupré and Gauthier Vernier) inserted 15,000 portraits dating from the 14^th to the 20^th century into the software, enabling it to make portraits itself. This led to a series of eleven unique images portraying the ‘Belamy family’, of which Edmond de Belamy is one.

The sale of Edmond de Belamy raises fundamental questions as to how centuries of IP law will interact with AI, the obvious question being whether there can be copyright in an AI-created portrait. Visually these paintings can hardly (if at all) be distinguished from paintings of human origin, supporting the view that if copyright protects expression (as opposed to an idea), there is an artistic expression to some extent. On the one hand, the AI can be seen as a tool used by a human artist. On the other hand, one could argue that the AI itself is emulating creativity, not the person feeding data to the machine. But is that enough, or even relevant for copyright law's purposes? Even assuming there can be copyright protection, a fierce debate can be held over who is the author and, therefore, who owns it. Is it Pierre Fautrel, who used the software to create the series of paintings and selected the paintings for the dataset, or are we in the realms of joint ownership with the creator of the algorithms? These questions will become more difficult to answer with the increasing potential of independently operating AI systems [particularly if AI datasets are pulling from open source material, notes the AmeriKat].

The answers to these questions can have far reaching implications, from everyday issues such as whether Edmond de Belamy can be freely copied by others, to more peculiar questions, for example, with respect to the droit de suite. With the fast-paced development of AI algorithms, their broad applicability, and the recent high-value auction of Edmond de Belamy, these questions will have to be answered. National courts are likely to grapple with these questions sooner rather than later.

Under Dutch law there is a rather low threshold for obtaining copyright protection. Hence, if presented in front of a Dutch court, copyright protection would likely be afforded as long as human choices – such as feeding specific information to the software – are involved. Even if no human choices are involved other than in creating the software, for example by connecting the software to Google images for its data input,  copyright protection would likely be afforded to the person who wrote the software, both for the software and the resulting painting.

A distinct IP question is what are people paying for when buying AI paintings? In legal terms they are buying a physical object. To the extent that any copyright exists (which may differ from territory to territory) they may also be getting an implicit or implied license. However, not all purchase agreements provide much clarity on what rights the buyer is getting. It is questionable whether amounts in the range of $432,500 would be paid for non-exclusive AI paintings. [Which creates a fun question as to whether, when Christie's delivers your painting, they also give you a hard drive with the code, says the AmeriKat].

A more policy driven question also needs to be answered:  Why are buyers willing to pay so much for AI art? Is it the concept of buying a new – and currently still rather unique – form of art?  In other words, is this a ‘first time value’? Or are people paying for the actual (artistic?) expression of the algorithms?"

Perhaps the answer to that last question may answer whether copyright subsists and should protect such works and which part of the subject-matter, muses the AmeriKat. If emotion has little to do with so many protectable works now, then why treat AI any differently?   Does it matter that this form of AI algorithm is using data sets to imitate previous work in creating a new work (i.e. GAN), as opposed to the CAN (Creative Adversarial Networks) system at Rutgers which is using data sets to create novel works - something different than the data?  Or are different AI systems just doing what artists have done for centuries - imitating and breaking the mold? 

On the final question of authorship, Caselles-Dupre explained that:

"If the artist is the one that creates the image, then that would be the machine. If the artist is the one that holds the vision and wants to share the message, then that would be us.  "

What do readers think?