AI Generated Art

by Danica Barreau
originally published in the American Society of Photographers Magazine and the VPPA Flash Magazine

Was Artificial Intelligence (AI) part of your vocabulary prior to 2022? AI’s been around for decades, and you’ve been using it every day. Have you used Google Maps or Waze to get to a destination recently? AI. Used your face to unlock your phone? AI. Asked Siri or Alexa a question? AI. Autocorrect, spellcheck, Google search, customer service chatbots, texts about weird charges on your credit card, photographing a check to deposit, Amazon and Netflix recommendations, auto-masking in Photoshop… that’s all powered by AI.

These applications of AI technology haven’t raised much concern or controversy. But professional photographers and artists worldwide are in an uproar and asking questions about an AI usage that applies to nearly everything we do. AI Art.

Exactly what is AI art? Is it legal, copyrightable, ethical? Should you use it? Will AI art kill fields like professional photography, graphic design, and illustration? By exploring the history, mechanisms, and current legalities of AI art, perhaps we can reach some conclusions.

 

What is AI Art?

The definition of AI Art is straightforward: artwork that was created with the help of artificial intelligence. AI artwork isn’t new, either. The first computer assisted digital art was created back in 1973 when a scientist and artist named Harold Cohen created the first ever AI painting using a program he named AARON.

The more current versions of computer-generated art rely less on a programmer directing tasks to create a piece of art, and more on machine deep learning technologies that allow the computer more autonomy in producing the images. In 2014, generative adversarial networks (GANs) were developed. GANs are the underpinning of generative AIs such as Midjourney, Stable Diffusion, and DALL-E 2. These AIs create images from text descriptions or “prompts”. Since they process the data in different manners, the overall style of the results differs between programs. While Midjourney can create striking painterly environments and characters, Stable Diffusion generates beautiful landscapes that are, in some cases, indistinguishable from actual photography. Many of the programs can also take an existing image as input to produce a creative variation of it, like those AI-altered selfies that popped up on social media recently.

To create AI art, algorithms aren’t written to follow a set of rules, but rather to learn a specific visual aesthetic by analyzing existing images to create something new. The AI is trained on massive datasets of images tagged with descriptive words like “orange”, “cinematic lighting”, or “Rembrandt”. The billion-dollar question is whether, when prompted to “paint an orange”, the AI made the image from pieces of images from the original training data or did it really learn what “orange” means. After viewing 200 images of Rembrandt paintings, does it really understand the artist’s favorite palette and tonal choices? To AI creators, the question seems ridiculous. The whole point of machine learning is to create flexible neural networks that learn like humans. Calling them collage makers defeats the entire purpose of the technology. But that question is at the heart of the AI art theft  and copyright debates.

 

Is AI Art Theft?

To build LAION-5B, the training set used by several of the AI image generators, bots crawled billions of websites, including large repositories of artwork at Getty Images, Flickr, Pinterest, and more. This is the same sort of internet crawl used by Google daily to power its search engines. Those images were then tagged with descriptive text to create word associations. LAION collected a library of 5.85 billion images*, including millions of copyrighted images gained without permission.

Several living photographers and artists have been dismayed to find out their names are being used to generate AI images. A website called “Have I Been Trained” (https://haveibeentrained.com) was made in response. Users can search for keywords that might bring up their work or they can upload a photograph they have taken to see if has been used. The group that created the website, Spawning, is also working on new tools for artist and photographer ownership of training data, allowing them to opt out of the future datasets used for training of AI models (but not the current one). Artstation added a #noAI tag to their images meant to keep the data crawlers for AIs off their images.

While all of that brings up a whole host of ethical dilemmas, it’s not illegal. There are several ways that people can use a copyrighted work without permission. That’s by design, to promote innovation and creativity. The fair use doctrine promotes freedom of expression by permitting the unlicensed use of copyright-protected works in certain circumstances. Case law holds that this type of use – ingesting code or content without permission to create new tools – is acceptable. Such as, say, training an AI to make art.

A recent class-action lawsuit filed by Joseph Saveri Law Firm argues that AIs are creating nothing more than complex collages and compete in the marketplace with the original images. They specifically call out prompts that use “in the style of” to create artwork that looks remarkably like something the original artists could have created and are requesting compensation for the named artists. Is it infringement to copy an artist’s style? Probably not.

While fair use doctrine does consider economic impacts, it focuses on whether the infringing use denies the original creator the market for their specific works. It doesn’t care if the new works diminish the overall market for all artists. It also doesn’t care if the new work creates unanticipated or previously unimagined value that’s not shared with the original creators. So many of these activities will probably be interpreted to be completely legal and permissible under current laws and precedents—part of the trade-off of copyright and public exceptions to “promote the Progress of Science and useful Arts,” as described in Article I of the US Constitution.

What is less clear is who, if anyone, will own the outputs of generative AI. This is an important question as, not to sound too dramatic, it could define the future of creativity.

 

What about AI Copyright?

Photographers are especially sensitive to copyright issues as they must, on an almost daily basis, remind clients of what they may and may not do with their images. In February of 2023, the US Copyright Office ruled for the first time against granting protection for artwork produced using an AI tool. The board found the AI-created image “lacks the human authorship necessary to support a copyright claim.” US copyright law doesn’t explicitly outline rules for non-humans, but case precedent has led courts to be consistent in the rulings that non-human creations (whether AI or monkeys) are ineligible for copyright protection.

Ahmed Elgammal, the founder of the Art and Artificial Intelligence Laboratory, developed AICAN, an autonomous AI artist. He fed the algorithm over 80,000 images representing western art over 500 years, with no particular focus on style or genre. Then he requested that AICAN create a piece of art, entirely on its own. The results were so good that AICAN pieces are exhibited worldwide and sell for thousands at auction. People genuinely like the artwork, and most can’t distinguish it from human generated pieces. Elgammal says that although he created the original algorithm, he has no control over what AICAN creates so he gives credit to the art pieces entirely to AICAN. It has been noted, though, that at a recent exhibit of AICAN work, both Elgammal and AICAN were credited with the work - perhaps to provide a legal bulwark in case copyright law is challenged.

Harold Cohen, the creator of that first computer-assisted art in 1973, has had his work exhibited in major museums around the world, such as the Victoria & Albert Museum and the Tate Gallery. A piece entitled “Secret” was sold at Christie’s for $11,382 in 2010 and, if you want to own a piece of history, there’s a 22x30 drawing by Cohen’s AI for sale at Aleator Press today for $8,500. All the artwork is clearly attributed to Harold Cohen because he was both the artist and programmer, providing his AI with detailed rules and forms (objects, plants, people) that allowed the program to create art. Cohen therefore owns the copyright on all the work.

Let’s consider this analogy. Imagine that you’re an author and you’re working with an illustrator to develop the cover artwork for your new novel. You provide descriptions and perhaps some sketches of the cover. You go through various iterations and finally decide on the final layout after which you’re provided with the finished artwork. You may have provided all the input, but it’s the illustrator that holds the copyright on it since they generated it. It's the same thing with the AI generator – it may be your unique concept and vision but since you didn’t generate it yourself, you can’t own the copyright. Since an AI isn’t legally a person, that means that AI generated artwork is not copyrighted. Although you could personally transform the artwork and then own the copyright under the Fair Act Index.

All AIs aren’t created equal. There’s “weak” AI, where the algorithm is designed to perform a specific task and can’t make any autonomous decisions. And there’s “strong” AI, where the programmer doesn’t have direct control over the final output. Art created by a weak AI could conceivable be granted copyright because it was merely a tool and remained under the direct control over the human making the creative choices. Strong AI, on the other hand, where the programmer doesn’t have direct control of the creative process, has less of a chance at copyright protection. The algorithm-generated images are purely arbitrary and mostly unpredictable.

In the process of prompting AIs for images, there is a creative input from the human, refining the output and training the AI in the process to get to the desired result. A human has some control of the creative input and although the results are not always predictable, they can be guided in a favorable direction. AI prompters can spend days developing the images to get to the point where they match what they saw in their mind’s eye, refining prompts and using previously generated images as weighted bases, called “seeds”, for new ones. There are open job listings today for “prompt engineers” because there is a science and art to understanding how to properly craft a text prompt to generate a specific image. This lands these new generative AIs in a grey area between “weak” and “strong” AIs. This is important in terms of copyright.

According to the World Intellectual Property Organization (WIPO), the world currently has two legal options:

The first, as demonstrated by the U.S. Copyright Office’s decisions, is to deny copyright to all non-human-generated content. Apart from the U.S., authorities in Australia and the European Union have settled similar cases by rejecting copyright applications on the grounds of works not being entirely made by human hands.

The second is to credit the creator for any work generated by any AI programs. This option is evident in the United Kingdom, as stated in Section 9(3) of the Copyright, Designs and Patents Act 1988, which not only gives credit to the human creator but also grants the work copyright protection. Other countries that have taken this approach include India, Ireland, and New Zealand. It will be interesting to watch these countries to see what sort of impact these decisions have on artist communities and their economic markets.

 

Is AI Art Really Art?

Ahmed Elgammal, AICAN’s creator, said: “I often compare AI art to photography. When photography was first invented in the early 19th century, it wasn’t considered art—after all, a machine was doing much of the work. The tastemakers resisted, but eventually relented: A century later, photography became an established fine art genre. Today, photographs are exhibited in museums and auctioned off at astronomical prices. I have no doubt that art produced by artificial intelligence will go down the same path.”

The confusion over who made the artwork, the machine or its user, complicates everything, and fuels arguments against recognizing AI generated artwork as a marketable artistic product. While the industry bickers about the definition of art, AI users are happily creating works and selling them, sometimes for sums they’d never get with traditional artwork. Because no matter how it’s created, it doesn’t diminish the aesthetic merit of the AI-generated piece.

If a human is directing the results and shaping them to fit their vision and to create an emotional response of their choosing, I believe it is art. Now, take the human out of the equation and leave the art generation purely up the machine, as with AICAN, you’ll have a harder time convincing me. While the AI can provide imagery that inspires, horrifies, or intrigues, it’s not creating art out of its own complex subconsciousness. It resonates because it’s expressing something it copied from its data set of human created images but it isn’t actually creating anything out of its own collection of emotions or experiences.

 

What About AI Photography?

There’s no such thing as AI Photography. The definition of photography is: “… the art, application, and practice of creating durable images by recording light, either electronically by means of an image sensor, or chemically by means of a light-sensitive material such as photographic film.” By its very definition, an AI version cannot exist.

A painting or illustration doesn’t need to pretend to be a photograph to have value. If you must produce an AI illustration to depict something you weren’t able to photograph, then you create art. But you don’t pass it off as photography. There is something viscerally disturbing about representing a moment, person, or place in time with a computer-generated image and calling it photography. A photograph has the trust of the viewer. It captured a moment exactly as it was and while you can enhance it with dodging and burning, split toning, and masking, that original capture, film or RAW, still exists. There is a “proof of life”. Photography is never going to be replaced by GAN-generated imagery because, at its heart, it captures something that actually exists. There is a tangible human-based connection that is missing when created by AI.

Most of the users creating AI photography are not actually photographers. At the time of the writing of this article, there were close to 115,000 posts on Instagram tagged with #aiphotography, ranging the gamut from nostalgia to future fashion concepts to images of real people in situations that never existed. That last category is going to be a problem for many. While generators of AI photography say they’re not infringing on any copyrights when they create people who don’t actually exist and don’t use the style of specific individuals or movies, the ones altering real people may find themselves in an unenviable legal position. 

Portrait photography centers around the entire experience of a session and an end-to-end solution, not just the final image. While there may be some who will use AI to create self-portraits, these are the same people who will pay a photographer $25 for 100 digital files. AI “photography” will probably decimate the market for low-end clients.  

While product photography seems endangered, the photographers I’ve questioned say that the stringent demands most of their clients have in terms of size and angle and lighting of their products make any true threat to their livelihood a distant concern.

Photojournalism will need to guard itself against false images – that is a serious risk and could affect world politics and national security. The potential for high-quality disinformation campaigns with falsified actors was already high – generative AI makes it a certainty. The identification of AI-generated images should be a high priority to any forward-thinking person.

 

How Do We Identify AI Art?

When AI generated images started gaining mainstream popularity in 2022, they were fairly easy to identify. The algorithms were still being updated/learning and giraffes would have six legs, eyes would have pupils looking in opposite directions, and the bane of artists everywhere, hands, seemed to stump the AI as well. There was a certain “look” to AI generated images that made them identifiable. But there is so much training data being provided to the GANs from millions of users playing around in programs such as MidJourney and Dall-E 2 that the AIs are learning, and fast. The improvement in renderings in just a few months is absolutely astounding.

Photographer Jos Avery found fame recently with striking black and white portraits posted on his Instagram account. It turns out they were created in MidJourney and “cleaned up” in Photoshop even though he originally claimed he shot them with his Nikon.

In a twist, digital Illustrator Ben Moran found himself banned from Reddit and his artwork removed because it was deemed to be AI-generated. When he offered to share the layered PSD file showing his work, he was told by one user to “find a different style” because it looked too much like AI.

Tidio ran an experiment to see if over 20,000 Internet users could recognize photos, artwork, music, and texts created by AI. In some survey groups, as many as 87% of respondents mistook at AI-generated image for a real photograph of a person. They did note that younger users and users familiar with neural networks tended to do better.

French photography magazine Réponses Photo recently published an issue with an AI photograph on the cover. The article inside was about the difficulty in identifying AI-generated images and how easy it is for people to be fooled. We urgently need a way to tell the difference.

Several companies are creating “GAN detectors” to try to identify GAN-generated images. Unfortunately, most are not particularly accurate. Microsoft recently released deep-fake detecting software called “Microsoft Video Authenticator” but only released it to news and political organizations. Microsoft has proposed a system where content creators can tag videos with a unique “signature” and if there are any edits to the video, it will generate a different signature, thus identifying a deep fake. That’s great for videos but what about photography?

To be cheeky, I posed the question to ChatGPT, a text-based AI.

What would be a good way to identify GAN-generated images?

… here are some methods that can be used to identify GAN-generated images:

1.       Expert analysis: Experts with domain knowledge or visual expertise can often identify GAN-generated images by examining the image's content, color, texture, and other factors.

2.       Metadata analysis: GAN-generated images may have different metadata than real images, such as inconsistent or missing timestamps, camera models, or location data.

3.       Statistical analysis: GAN-generated images may have different statistical properties than real images, such as differences in pixel distribution or correlations between pixel values.

4.       Deep learning: Deep learning models can be trained to identify GAN-generated images by learning the statistical differences between real and generated images. For example, a classifier can be trained on a dataset of real images and then tested on a dataset of mixed real and GAN-generated images. If the classifier can accurately distinguish between the two types of images, then it is likely that the image is GAN-generated.

5.       Reverse image search: Reverse image search engines can sometimes identify GAN-generated images by comparing them to other images on the internet. If the search engine cannot find any matches, it is possible that the image is GAN-generated.

The answer seems to be “good luck”. For things such as image competitions, a combination of expert and metadata analysis may be the solution. Perhaps uploading a RAW of the original shot that includes metadata. But for the moment, we’re going to have to rely on our own critical eye and keep abreast of new developments in GAN technology. And while there are currently some generalizations that can be used to identify AI-generated portraits, give it a few years and those will no longer apply.

 

In Conclusion

At this point, there’s no going back. The time to act was a decade ago, before the developers of generative AI systems started crawling the internet to train the algorithms.

The idea that there was a theft committed and that artists must have compensation doesn’t work when you think it through. Individual artworks don’t have any value once they’re broken down into an algorithm. At best, a collective value could be assigned and, once divided between the 5.85 billion works integrated into LAION, would be miniscule. And that’s only if copyright law is shown to be violated in the first place.

AI generated artwork shouldn’t be copyrightable. It doesn’t mean that the work isn’t valuable or couldn’t be sold. But it does mean that everything AI makes would immediately be public domain and be available to all the other makers to use as part of their own creative process. If it sounds unfair, remember that it’s exactly what the AI creators did when they scraped the internet to train their algorithms.

Another argument to not allow copyright protection is the need to prevent the suffocation of innovation. Every new computer copyright would immediately be followed by automated lawsuits to defend them. Would there be any room left for creation, progress, or cultural development? AIs might offer important opportunities in terms of art but making them a legal monopoly is not the answer.

AI technology is evolving at fantastic rates and the creative industries will need to evolve with it. The copyright issue for new works must be addressed and the identification of AI-generated artwork should be everyone’s top priority, especially photographers. It’s important for the creators of AI-generated art to be transparent about the process they used to create the artwork and to credit any sources that were used in the creation of the finished product. And it’s incumbent on us, as photographers, to stay abreast of the technology so that we can identify false images and help find solutions.

Much of the backlash from traditional artists comes from seeing results generated in minutes that would have taken them years of learning and hours of labor to achieve. Not only is their livelihood threatened, but it was their work or those of their teachers and predecessors that was mined to achieve these images and they have not been compensated. And while I can well understand their anger and frustration, it’s just not possible to put the genie back into the bottle. We will all have to learn to adapt and find ways to make the technology work for us. That’s not to say we shouldn’t question or voice concerns – that’s the only way we are going to find solutions to these new AI issues.

While I don’t foresee the extinction of photography, those in a race to the bottom will not be able to compete with AI. Photographers will need to continue to focus on creating a unique experience for their clients and explaining the value of their art. We need to figure out how to protect and preserve human art in a way that acknowledges it as a public good. Otherwise, art will have to compete economically with tireless machines capable of endlessly churning out soulless imitations.

 

*Technically LAION collected the URLs for images, not the images themselves. An important distinction being used by LAION developers to cede any responsibility for copyright infringement to those using the data rather than themselves.

All images herein were generated in MidJourney v4 using text prompts.

 

  

References

 Rockwell, A. (2017, August 28). The history of artificial intelligence. Harvard University. https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/

Staff, A. (2022, July 5). What is the first AI art and when was it created? NightCafé. https://nightcafe.studio/blogs/info/what-is-the-first-ai-art-and-when-was-it-created

Garcia, C. (2016, August 23). Harold Cohen and AARON – A 40-year collaboration. Computer History Museum. https://computerhistory.org/blog/harold-cohen-and-aaron-a-40-year-collaboration/

Ipstein, Z., Levine, S., Rand, D., Rahwan, I. (2020, September 25). Who gets credit for AI-generated Art? ScienceDirect https://www.sciencedirect.com/science/article/pii/S2589004220307070

Elgammal, A. (2019, January) AI is blurring the definition of artist. American Scientist https://www.americanscientist.org/article/ai-is-blurring-the-definition-of-artist

Whiddington, R. (2023, February 24). A photographer who found Instagram fame for his striking portraits has confessed his images were actually AI-generated. Artnet. https://news.artnet.com/art-world/fake-instagram-photography-ai-generated-joe-avery-2260674?fbclid=IwAR250jwS5Vhqj09rXIEyJUhK0ektJjQ8JrC8zNGpIEwJkyFy79rtIKrmaVY

Whiddington, R. (2022, December 16). Independent artists are fighting back against AI image generators with innovative online protests. Artnet. https://news.artnet.com/art-world/independent-artists-are-fighting-back-against-a-i-image-generators-with-innovative-online-protests-2231334

Recker, J. (2022, March 24). US copyright office rules ai art can’t be copyrighted. Smithsonian Magazine. https://www.smithsonianmag.com/smart-news/us-copyright-office-rules-ai-art-cant-be-copyrighted-180979808/

Jeong, S. (2018, April 24). Appeals court blasts PETA for using selfie monkey as “an unwitting pawn”. The Verge. https://www.theverge.com/2018/4/24/17271410/monkey-selfie-naruto-slater-copyright-peta

Rajnerowicz, K. (2022, August 23). Human vs AI test: Can we tell the difference anymore? Tidio. https://www.tidio.com/blog/ai-test/

Growcoot, M. (2022, September 19). You can now check if your photos were used to train AI image generators. PetaPixel. https://petapixel.com/2022/09/19/you-can-now-check-if-your-photos-were-used-to-train-ai-image-generators/

Growcoot, M. (2022, August 19). AI image generators could be the next frontier of photo copyright theft. PetaPixel. https://petapixel.com/2022/08/19/ai-image-generators-could-be-the-next-frontier-of-photo-copyright-theft/

Mishkin, P., Ahmad, L. (2022, April). DALL-E 2 preview – Risks and limitations. Github. https://github.com/openai/dalle-2-preview/blob/main/system-card.md#copyright-and-trademarks

Chambers, SW. (2020, December 22). AI agents are not artists. Towards Data Science. https://towardsdatascience.com/artificial-intelligence-agents-are-not-artists-9743d5dba2d0

Di Donfrancesco, G. (2023, January 6). Is AI photography as ethically plagued as AI art? Daily Dot. https://www.dailydot.com/debug/ai-photography/

King, S. (2023, February 13). Why film photography is the antithesis of AI art. PetaPixel. https://petapixel.com/2023/02/12/why-film-photography-is-the-antithesis-of-ai-art/

Foley, J. (2023, February 9). Photo magazine’s AI-generated cover just sent shivers down my spine. Creative Bloq. https://www.creativebloq.com/news/ai-photo-magazine-cover

Merkley, R. (2023, February 28). On AI-generated works, artists, and intellectual property. Lawfare. https://www.lawfareblog.com/ai-generated-works-artists-and-intellectual-property

Kovacevic, A. (2022, November 18). Intellectual property law becomes murkier as more creations are created with AI. Smart Data Collective. https://www.smartdatacollective.com/intellectual-property-law-becomes-murkier-as-more-creations-are-created-with-ai/

Nedee, I., Kasteleijn, L. (2021, April). Artworks made by computers: (im)possibility of and (un)desirability of copyright protection of AI made. Art Law Services. https://www.artlawservices.com/post/artworks-made-by-computers-im-possibility-and-un-desirability-of-copyright-protection-of-ai-made