If you are not an early tech adopter, you may not have heard of, or heard much about, DALL-E 2. Not being on the cutting edge of hi-tech myself this term had not crept into my consciousness until a couple of weeks ago, while I was up at my cottage north of Toronto enjoying the autumn foliage. While basking on the dock in my “Muskoka chair” (aka Adirondack Chair for my US friends) noodling on my phone, I saw reports that Microsoft will be soon be bringing Dall-E 2 to a laptop near you by integrating it with Bing and Microsoft Edge. Dall-E 2 is an artificial intelligence (AI) program that produces artwork through the use of verbal and textual prompts. Sort of like, “make me a picture of cats and dogs wearing funny hats in the style of Rembrandt”. Or to be more contemporary, “make me a picture of purple dragons breathing fire in the style of Greg Rutkowski”. (More on Rutkowski later).
As I have written about in the past, the expanding use of AI raises many issues for creators and for copyright; see, for example, my blog post on “AI, Ethics and Copyright”. These range from ownership/authorship to liability for infringement to dealing with piracy, both of and by AI-generated content. Who is the author of an AI generated work? The US Copyright Office (USCO) has taken a clear position that there must be a human hand behind an AI-generated work for it to be subject to copyright. As I wrote earlier this year, the USCO rejected on appeal the application for copyright registration of an art work that the applicant claimed was the exclusive creation of a machine (one that he had invented). In the UK, by contrast, there is a provision for copyright protection for machine-generated works where no human author of the work is involved. The term of protection for such works is less than that of works authored by a human.
The ease of producing a passable image through voice or text prompts raises once again the question of who owns the copyright on an AI-generated work, or even whether a work like this can be subject to copyright protection? If I had exercised some skill and originality, mimicking the style of Monet which I had studied from artbooks, online images or the real paintings while looking at the foliage across the lake from my dock, arguably the copyright in the artwork would be mine. But all I did was type in a few words, allowing the algorithm to compile four images from countless impressions of foliage, chairs, docks and Monet paintings scraped from the internet, of which I chose one. The only originality I exercised was the choice of prompts. Moreover, the identical prompts produce slightly different images each time, so there is a degree of randomness introduced into the process by the algorithm. While it is generally accepted that a photograph can be copyrighted, even one taken by an amateur like myself, because in pointing the camera, choosing the location, judging the light etc, I have made some creative choices even though all the heavy lifting was done by my phone camera, even this minimal degree of creativity does not exist with text prompted art.
However, while issues of authorship will continue to arise and be debated, the most serious current copyright challenge arising from AI relates to the widespread, unauthorized text and data mining (TDM) of copyrighted works. Massive amounts of data are required to feed the AI machines that produce AI-generated artworks through DALL-E 2, developed by OpenAI, or other similar programs such as Midjourney, Meta’s Make-a-Scene and Stability.AI’s open-source AI art generator, Stable Diffusion. (You can try it here, but please don’t input the name of any contemporary artists or artists whose work is still under copyright. Monet—see image on this blog post–died in 1926 and his works are in the public domain).
What is the data the AI machines ingest and process to create a work like the one at the top of this blog post? It is the work of artists; past artists like Michelangelo, Rembrandt, Monet and Van Gogh, all in the public domain, or Picasso or Dali, whose works are partially or not in the public domain, or contemporary living artists with unique styles, like Greg Rutkowski and others. Some 6 billion images were scraped from art sites and other depositories on the internet to create the data base. Rutkowksi is a contemporary Polish artist whose work has appeared in a number of video games. His art-style is notable for fire breathing dragons on eerie futuristic landscapes. As outlined by an article in the MIT Technology Review, his distinctive style is now one of the most commonly used prompts for AI generated art, but this is not necessarily to his benefit. Type “Rutkowksi” into a search engine and you are as likely to get an AI-generated clone of Rutkowski’s work as the genuine article. The AI generated images threaten to swamp his original work and open the very real possibility of being in competition with the original work of the artist. This has led Rutkowski to proclaim that AI should not be scraping and using the work of living artists.
Rutkowski does not appear to be against AI-generated art, but that does not extend to allowing his art to become the feeding ground to train algorithms. Other artists also have an ambivalent attitude to AI “art by numbers”. CBC interviewed a number of artists who saw programs like DALL-E 2 as a “tool” that they could use, although they recognized its shortcomings and challenges. Perhaps one way of differentiating attitudes to AI art is whether one is an established artist, whose work is being used to help generate the program, or an artist who enjoys using the tool to create new art forms and get noticed. Certainly not all the inputs are drawn from copyrighted works. One solution would be to limit the AI crawler to works that are in the public domain.
What can artists do to prevent their works from being used? Because the images are scraped from the internet, one solution is to insert watermarks since some of the web crawlers are programmed not to accept watermarked images. That does not, however, get around the issue that in many instances the AI platform is still violating copyright with impunity. Artists should not have to request that their art be removed from the crawler’s inventory (in any case, this seems to be almost impossible to do), nor should they be required to take defensive measures, or legal action. Lawsuits are almost certainly going to arise. In response, Getty Images, one of the leading providers of “stock images”, (art, designs, photos), has announced that it will not carry any AI generated art. What Getty provides to its subscribers is certainty that the images they are using are legally licensed. Getty has realized that with AI-generated art it may not be able to provide this certainty.
With the application of technology to information distribution and art, we have seen how the law of unintended consequences can prevail. Algorithms can be a useful tool to sort information and offer content to users according to their preferences, but they can also result in reinforcing the biases and intolerances of users, resulting in a closed-circuit echo chamber of hate and conspiracy theories. We have all seen how online discourse, potentially a positive tool to promote free expression and a wider exchange of views, can degenerate and slide into harassment and much worse. In the same vein, AI generated art presents a myriad of challenges in terms of people misusing the technology to create fake or offensive images, to distort artworks, hijack trademarks or a person’s image, etc. While some attempts have been made by the AI art platforms to deal with these problems by installing filters, we all know how ineffective automated filters have been in other online environments. However, while there has been some acknowledgement by the AI platforms that the form of content needs to be regulated to avoid abuse, there has been no self-restraint whatsoever when it comes to freely helping themselves to content owned by others—content protected by copyright–to feed the algorithm which produces work that will likely negatively affect the income prospects of the rights-holder.
In my view, this is the greatest threat that AI generated art imposes—the unrestricted data mining of protected content to produce derivative works that pose a very real threat of competing with the work of the original artist. It behooves all those who treasure creativity, and the protection afforded it through copyright to sound the alarm and push back.
This emerging challenge is enabled by a lax and permissive attitude to text and data mining, all in the name of supporting “innovation”. In the US, the proponents of this unrestricted data mining of copyrighted works for commercial purposes argue that it is “fair use”. Given that the end product scans the entire original copyrighted work and competes with and has a substantial negative effect on its commercial exploitation, it is highly questionable whether uses such as DALL-E 2 and others are fair use. This will surely be tested in court. In other countries, where there is a clear legal definition as to what constitutes a user exception to copyright, there is either no exception for data mining or a very limited exception. However, there are pressures to throw the door more widely open, particularly in the UK, which would set a dangerous precedent. I will follow up on these developments in a subsequent blog.
This article was first published on Hugh Stephens Blog