Generative Art and Copyright — Part I

By Felix Yuen

When art meets new technology, it often raises the question whether the existing intellectual property regime is still suitable. This is the case for AI-derived art, and also a related but distinct art form known as generative art. The first part of the two-part series gives an introduction to this art form.

What is generative art?

There are more technical, legal definitions out there but I find this one given by the art museum Tate Modern in the UK short but succinct:

‘Generative art is art made using a predetermined system that often includes an element of chance — [which] is usually applied to computer based art.’

This definition highlights an important key point of generative art. It is a combination of the creator’s artistic choice and a certain degree of randomness pursued by the system. The creator sets the framework for the apparatus to run, and the apparatus creates random results within the framework. It therefore differs from other forms of art in that there is an element of uncertainty and relinquishing of control by the creators.

The root of generative art may date back to the Middle Paleolithic era, but with the invention and development of computers in the 20th century, generative art became common for an obvious reason — computers and the programs that run on them are now much more accessible and user-friendly. Vera Molnár has been described as the pioneer of modern generative art who started using computers to create algorithmic visual art pieces in the 1960s.

One of Vera Molnár’s earlier computer-generated works, ‘Molndrian’, 1974 (image reproduced for review, criticism or otherwise fair use)

As more and more people with a technical computer background enter the generative art scene, creators no longer need to design the system by themselves from scratch. What they can do now is picking the suitable and readily available programming codes and software designed for this purpose, and adjusting the parameters of the codes and software to achieve a desired outcome. The computers will then do the rest and generate results in a split of a second.

After 2014 when the Generative Adversarial Networks (GANs) technology was introduced to the world, this artificial intelligence model became the popular medium for delivering generative art. Some technologists have done an excellent job in explaining how GANs work (for example, here). In short and in layman terms, GANs involve two components — a generator that produces new data that resembles the data that is fed to the machine (the training data), and a discriminator that gives feedback as to whether the data produced is ‘real’ training data or ‘fake’ data produced by the generator. The generator learns to improve in the next round, and the loop goes on to the point where the ‘fakes’ and ‘reals’ are indistinguishable. GANs are said to be an important development in the area of machine learning and are found to be very effective in producing realistic visual images of non-existent persons, for example, which even humans are unable to tell apart.

A GAN research project in 2017 called ‘GANGogh’ that creates Van Gogh-styled drawings (image reproduced for review, criticism or otherwise fair use)

Generative artists use GANs to produce artworks that look extremely natural as if they are works of humans of a particular style. Technologies however do not stop there and newer, more powerful deep learning tools have been developed which are said to be superior than GAN in creating realistic images. For example, the AI research lab OpenAI recently announced DALL-E 2 that turns natural language captions to images using CLIP and diffusion deep learning techniques. It can also make realistic edits to existing images based on natural language, or create different variations inspired by the original.

An image created by DALL-E 2 that is resulted from the caption ‘a painting of a fox sitting in a field at sunrise in the style of Claude Monet’ (image reproduced for review, criticism or otherwise fair use)

Like other digital art, generative art pieces have been sold as NFTs. One obvious reason is the marketing effect that the combination of new techs can bring. Another reason is that creators can easily generate hundreds and thousands of works by re-running the programs on the same parameters. The resulting works are each unique but have the same ‘style’ that fit the criteria of NFT collections. A well-known example is the series of ‘Autoglyphs’ art developed by LarvaLabs, the creators of the CryptoKitties NFTs.

Generative art ‘Autoglyphs’ that have been sold and resold for millions of dollars (image reproduced for review, criticism or otherwise fair use)

Generative art made by these AI models is able to reach beyond human’s own originality or imagination, which is exactly why it creates a problem in copyright. We will look into this issue in further details in the second part of this series.

The res ed cohort programme is an initiative by resolutio to help spread awareness on NFT rights. Learn about our cohort here. For updates, follow us on Twitter & LinkedIn and join our Discord Community.