AI paintings, questions of ownership, and the Internet of Art Things

The world art market has been thrown into some turbulence by the record $432,500 paid at auction for a portrait generated by a machine learning algorithm. It has also provoked controversy within the AI community over ownership, attribution and originality, although these concerns do look like red herrings. True to form in the AI trend, the painting was heavily oversold amid media hype and as a result beat the auction house estimated range of $7,000 to $10,000 by about 45x.

The portrait of a fictional character from antiquity, Edmond de Belamy, was created by Obvious, an AI studio in Paris, run by three 25-year-old researchers Hugo Caselles-Dupré, Pierre Fautrel, and Gauthier Vernier. It uses a type of neural network-based machine learning algorithm sometimes referred to as a generative adversarial network (GAN).

This was developed in game theory, through combining two neural networks. The first is generating candidate models from original data sets. The second then discriminates between a new model and the original data. The machine learning exercise is then to refine the models by repeated iteration, until they match the form of the original data sufficiently closely that the second discriminating neural network fails to distinguish them as being different.

In this case the GAN was fed a data set generated from 15,000 portraits, painted between the 14th and 20th centuries. The learning run generated images from those and refined them, until they could not be distinguished from the original training set. The model was then applied to generate a novel portrait with some deliberate distortion which blurs the face to make the image appear unique rather than looking like a copy. The image was effectively an amalgam of numerous original portraits from the data set.

The hype that led to the huge excess in price paid over the estimate was seeded by a well-judged article in Christie’s online magazine, which then went viral with plenty of errors in replication such as the false claims that it was the first portrait generated by AI and was made by the machine with no human artist involved.

There have in fact been plenty of art works made by some form of AI algorithm, dating back to the 1970s and 80s when UK artist Harold Cohen was the only person to achieve significant success producing images from a rule based expert system, some of which were displayed by the Tate Gallery in London. Although Cohen himself continued to fly the flag for AI-based art almost until his death in 2016, the field otherwise stayed quite dormant until the recent surge in AI on the back of contemporary machine resources, attracting a number of artists and some auction sales.

Among artists exploring the process now are Tom White, Mario Klingemann, Anna Ridler and Robbie Barrat. It was in fact Barrat who wrote the original code on which the Obvious algorithm is based, as well as developing some of the dataset. Obvious has admitted that the code has not been greatly modified but it is not clear whether Barrat will receive any of the unexpected windfall resulting from the sale.

The issue of attribution then comes in if we consider this form of AI-assisted art as conceptual and residing in the process. Great artists have always stolen from their predecessors to obtain and assimilate new ideas and concepts. Works are considered original when they are transformations rather than just imitations of earlier works. The same criteria could be applied to algorithms and in this case the work should be considered original if the underlying process is sufficiently advanced on Barrat’s algorithm.

It is likely in any case that this form of AI-generated art may end up being considered too derivative, unless and until substantial advances in the transformation are achieved, for Edmond de Belamy is not a very good painting. It fetched such a high price only because the buyer judged that as a result of the hype it would be famous for being famous and go down in art history as a landmark that would likely increase further in value.

The more interesting creations are coming not so much from advanced algorithms but developments around the Internet of Things, leading inevitably to the designation Internet of Art Things (IoAT). The focus is more on 3D installations fed by IoT devices than traditional 2D paintings.

One of the most celebrated examples to date is the eCLOUD Project, an installation at San Jose International Airport in California. The construction of polycarbonate tiles is not itself so noteworthy, but the tiles are configured to switch gradually between transparent and opaque states controlled by real time weather data from NOAA locations around the world.

The data generates a simulation representing weather from any of the international locations by turning the individual tiles on and off in a certain pattern. The simulation is displayed within the cloud sculpture, as well as on a dynamic display at eye level in the airport terminal. This impressive display has spawned other such large-scale projects fed by IoT data in different contexts.

Of greater interest perhaps for the IoT world itself is work by more technical artists looking at cross disciplinary applications where the artworks themselves become connected and function as IoAT devices. Among promoters of the IoAT in this context is Sean Clark, a UK artist and curator, who is also founder and managing director of web/mobile developer Cuttlefish, as well as director of Leicester based arts company Interact Digital Arts.

Clark argues that the IoT and AI should be seen not as digital substitutes for the creative process but new tools to enhance it, so that art continues to be made by human hands and brains. He has been working in what he calls an IoAT infrastructure featuring a new message broker based on Message Queue Telemetry Transport (MQTT), the ISO publish and subscribe protocol. Messages, says Clark, can then be transmitted between artworks and associated software processes as connected things, spawning new ideas not just for collaboration but also installations and assemblies created by individuals.

These are early days and Clark does not spell out which components would be connected but does argue that MQTT will provide a major stimulus not just for the IoAT but the IoT in general, just as Apache did for the World Wide Web from 1996 onwards, driving it into the public domain. So it could be this Christie’s sale may turn out to be just a side show or curiosity before the IoT and AI combine to achieve a much more radical transformation more in harmony with the human creative process.