Art Creation: AI Tools, Copyright Law and the Explosion in AI Art Explained
Exploring the Legal and Ethical Implications of AI-Generated Art and Copyright Ownership in the Digital Age
As everything interesting in AI and the law lately starts…with a lawsuit.
Midjourney and Stability AI, the creators of AI art generators Midjourney and Stable Diffusion, and the popular artist portfolio platform DeviantArt, which recently launched its own AI art generation tool named DreamUp, have been sued by three artists.
For the AI art copyright lawsuit, Matthew Butterick, a lawyer and typographer, has teamed up with the Joseph Saveri Law Firm to file on behalf of three artists. Both lawyers currently have a similar suit against Microsoft, GitHub, and OpenAI. Saveri handles antitrust and class action cases, and the copyright lawsuit is similar to the one filed by them relating to the AI programming model CoPilot. The basis of the CoPilot lawsuit is that CoPilot was trained using lines of code gathered from the internet. (A future article will analyze this case).
The copyright case argument is that these AI art tools enable users to create unique pieces of art by training the machine learning models on copyrighted data. Since the tools were trained on copyrighted works, the output of each of them, or perhaps the profits the companies earn from their use should be allocated to the artists whose works were consumed by the tool to enable it to create art.
Plaintiff’s counsel’s position is that every piece of art these tools create is infringing. To be clear, the allegation is not that the tools create exact duplicates of the plaintiffs’ works. It is more akin to the claim that the output of the AI art tools are so reliant upon the trained model (using the plaintiffs’ copyright art) that the output works are sufficiently linked to those copyrighted works to require compensation to the original artists.
AI Art in the Style of Your Favorite Artist
The three main tools mentioned above, Midjourney, Stable Diffusion and DreamUp art models trained on billions of images collected from the web, generally without the creators’ knowledge or consent. They are being used to create artwork that replicates the style of specific artists. Meaning, you can create endless works of new art that appear in every way to be in the style of Van Gogh’s Starry Night for example. The image below was created by me as I was writing this piece by entering this simple prompt into a free version of the Stable Diffusion tool for Mac called DiffusionBee. The prompt was “cityscape of manhattan in the style of van gogh's starry night”
However, not all the “style” prompts work as well. Below is the result of the slight modification of the above prompt into this: “cityscape of manhattan in the style of dali”
I am no art critic, but that is not what I think of when I hear “in the style of Dali.”
The artists in the lawsuit have, as one of their arguments, that their unique style is part of their copyrightable content that is now accessible to be duplicated in any piece of AI generated art.
A recent case involving a Disney illustrator presents some challenging trade offs in this area. Hollie Mengert is a Disney illustrator who discovered a Canadian mechanical engineering student used her artwork in an AI experiment. He used 32 of her pieces and spent a few hours training a machine learning model to replicate her style. She had been working since graduating art school in 2011 to develop her style which was now easily reproducible in an endless array of similarly appearing images.
This definitely raises questions about fairness but also about whether Mengert has any recourse in this situation and whether she should. What about the student? Does he own the copyright on the output of the model he created?
The copyright office recently granted a copyright to a comic book that was generated by AI using MidJourney. The key factual/legal issue on whether this registration will be invalidated if challenged is the degree of human effort or input. The extreme example of simply entering a text prompt into a program and then seeking to copyright the result is likely to fail. But, for a person who spends hours or days perfecting the now obviously intricate creation of text prompts, using various settings on the tool and perhaps even using photoshop or other media to modify the result further after the AI art tool completes its process, they may well receive copyright protection for such a work.
In a sense, can a painter paint without someone making canvas, brushes, the paint itself and other chemicals that are used to modify the paint? No. But, can a brush company claim copyright in works made with their tools? Also no. An artist using various tools, brushes, etc. can claim copyright in the resulting work. An argument can emerge here that AI art tools are merely advanced versions of brushes, canvas and paint. They are the tools to create something, but they are not part of the actual creating of the thing. And, the person entering the prompt is using those tools to create something novel. The entire “level of human involvement” question seems challenging to regulate given that I have seen pieces of art in galleries so intricate I imagine the artist worked for days or years to complete it. Likewise, I have seen pieces by other artists that are a single, uniformly applied color to a rectangular canvas which likely took five minutes or less from conception to completion.
Below is a work titled Blue Monochrome by Yves Klein which has hung in the Museum of Modern Art in New York for years. Ponder the amount of time to conjure and then physically complete this work. Then see below.
Below is a photograph I took of a work titled The Dream by Salvador Dali at the Cleveland Museum of Art.
There is certainly an argument that Mr. Klein’s work took significantly less effort to imagine and bring to life than Dali’s work. But, under current copyright law, they both retain protection.
With this reality, the input of a prompt in to a tool resulting in a novel image would appear to qualify as sufficient human involvement to bring copyright protection. But, alas, the courts, notoriously technophobic, will have to sort that out.
Copyright Implications for AI models that used copyrighted works for training
All of the AI art generating tools were trained on billions of images scraped from hundreds of websites. A recent analysis of that data determined that approximately 8% of those images were scraped from Pinterest alone. Other platforms like WordPress, BlogSpot, Flickr and DeviantArt were also used. Wondering if your image was used, you can check here.
The companies behind the AI art tools lead with the defense of fair use which the copyright office summarizes as:
Fair use is a legal doctrine that promotes freedom of expression by permitting the unlicensed use of copyright-protected works in certain circumstances. Section 107 of the Copyright Act provides the statutory framework for determining whether something is a fair use and identifies certain types of uses—such as criticism, comment, news reporting, teaching, scholarship, and research—as examples of activities that may qualify as fair use.
Most copyright experts agree that two factors are the most important when applying fair use principles:
The purpose or nature of the use and; (does it transform the original?)
The impact on the market. (does it potentially destroy the market for the original artist?)
Applying these two it seems most likely that the training of the models using the copyrighted works would be covered by fair use. However, the generating of new content via these tools appears to be outside that protection. An analogy might be if you trained a tool on 100 songs by the artist Ed Sheeran and then unleashed that tool to create new works in the style of Ed Sheeran. That AI tool output would be directly competing with Sheeran and likely raise copyright concerns.
We Did This With Napster
Some experts and others commenting on this debate say we have had this problem before and resolved it. The music industry was caught off guard by the popularity of the open source file sharing service called Napster. Some legal fights ensued, but the industry did not fold, nor abandon a digital only delivery of music. It evolved quite quickly into tools like Spotify and Apple Music which both deliver music in much the way Napster did, but also compensate (some would argue very poorly) the artists featured in those services. Perhaps AI art tools will arrive at the same place.
How It Will All End?
Anyone who confidently tells you how this will resolve is not to be listened to. This is a novel question involving interlocking concerns, statues, factors to be applied, etc. It will travel through lower courts and undoubtedly to appeals. The U.S. Supreme Court is currently considering a copyright case involving the use of a photograph of Prince by a photographer from decades ago which Andy Warhol incorporated into a work that he copyrighted and his foundation has earned millions from licensing. That is a case to watch for what the court may well think about the image use issues in these AI art cases that will arrive on its doorstep in the years to come.
One argument that came to mind when reviewing all this background material for this post was the way in which art evolved over the centuries before AI…like all the centuries.
Artists Influence Other Artists and Styles are Copied
Imagine a 22-year-old painter in 1875 who is enthralled with, let’s say, an artist named Jean-François Millet. Millet was at that time the most famous artist in Europe. He sold paintings for sums that would be in the millions today. This artist sees one of Millet’s exhibitions and is inspired. Six years later, that aspiring artist even drew copies (read that again, copies) of Millet’s existing prints to help improve his own drawing skills. By doing so, he was inarguably, acquiring Millet’s style.
As an example, this now more experienced artist, in 1889, painted something very much like a previous Millet work.
Millet’s work is on the left and the artist who painted the work on the right….Vincent Van Gogh.
The three artists suing the AI art tool makers are in a way arguing that what Van Gogh did back then if done by an artist today would trigger copyright issues for Van Gogh. Imagine an artist today simply staring at 15 different works by the same artist and then, creating their own unique work but “in the style” of those 15 works by that artist. Would that artist who composed those 15 works be entitled to copyright payments in the output of the later artist who used those works to mimic their style, but created a unique work? We will find out some day soon.