When AI takes on the role of artist

In 1901, American architect Frank Lloyd Wright gave a speech in Chicago on the role of the machine in the creative process. He explained that “its function [was] ultimately [to] "emancipate human expression." A century later, this “machine” takes the form of artificial intelligence (AI).

It is capable of creating any image from a textual description – text prompts – in a few seconds. Would you like to see the English naturalist Sir David Attenborough fighting a polar bear with his bare hands? Or American-Canadian actor Dwayne Johnson meeting his alter ego, The Rock? It's now possible thanks to software such as Midjourney, DALL-E 2, Imagen, DreamBooth and Stable Diffusion.

These tools use language understanding and learning models on huge quantities of data to generate images from a line of text. Google explains that it trained Imagen on LAION-400M, a database of 400 million images associated with written captions found on the internet.

This is where the artificial intelligence gets its "inspiration." It then generates several illustrations corresponding to a request or prompt – for example, a teapot with muscles – by reducing image noise in a cluster of pixels of random colors (denoising). It is this "diffusion" technique that makes the current image generating software so successful.

Since June, DALL-E 2, Midjourney and company have seen their popularity skyrocket on social networks, largely thanks to the sometimes realistic and sometimes grotesque visuals they produce. Some illustrate fairly vague concepts like blockchain or joy, while others respond to very specific requests (like "quails listening to angry parents at a school district board meeting").

Generative artificial intelligence software is capable of responding (more or less successfully) to any request sent by a user, as long as it doesn't go against its moderation policy. For instance, OpenAI, the company behind the DALL-E program, prohibits the generation of violent, pornographic or political images, and prevents the creation of "photorealistic" illustrations, especially of public figures. DALL-E would thus refuse to imagine a fight between Donald Trump and Vladimir Putin.

But that’s just about the only limit this technology places on "human expression," to use Frank Lloyd Wright's phrase. Programs like Imagen and Midjourney don't just make composites of images available on the web; they also create new ones. And they do this in a multitude of "styles": painting in the style of Picasso, Pop art, Impressionism, street art, even cyberpunk or Bauhaus.

While these generative artificial intelligence systems are rather easy to use, it takes some time to find which turn of phrase will give satisfactory results. Contrary to what one might think, typing in random keywords as a prompt is not enough to get a unanimously acclaimed work. "AI art generators never work exactly how you want them to. They often produce hideous results that can resemble distorted stock art, at best," digital artist Erik Carter told MIT Technology Review.

In particular, these programs have difficulty imagining anatomically correct animals or humans. Most of the faces they generate look "squashed," or at least distorted.

Philosopher and journalist Tom Whyman writes in ArtReview that AI-designed characters "appear as they do when you imagine them, not quite complete."

Another concern is that generative AI tools reproduce certain racial and gender stereotypes. DALL-E 2, Midjourney and others tend to represent caregivers as women and doctors as men. They also convey images based on a very Western canon, due to the lack of diversity in the millions of images they are trained with.

OpenAI has already announced various measures to ensure that DALL-E 2 is not an additional vector of discrimination on the internet. The company has the support of millions of internet users who, like the company, want to improve the images produced by its flagship program.

Hundreds of thousands of users gather on the DALL-E 2 Discord community servers to exchange tips on how to improve their text descriptions -- a process known as "prompt crafting" – or on how to overcome the limits of artificial intelligence. "Does anyone have any advice for getting Dall-E 2 to generate characters from popular TV shows?

I have no problem coming up with heroes like SpongeBob SquarePants, but for Patrick Star, for example, it generates a starfish, not the character himself," asks one user. Another turned to members of this online community for help in producing a centaur.

This phenomenon is not limited to DALL-E 2, or even Discord. Users of generative artificial intelligence software also convene with other enthusiasts on Reddit and Twitter to discuss the possibilities. Some aficionados, like Guy Parsons, have even published digital guides teaching readers how to best use these new tools of artistic creation, and maybe even one day create genuine masterpieces.

Last August, American artist Jason Allen even won an art contest organized within the Colorado State Fair with a painting generated by Midjourney's artificial intelligence. The work in question, entitled "Theatre D’Opera Spatial" or "Space Opera Theater," required 80 hours of work, or more precisely, 80 hours of "prompt crafting."

But when this prize -- however modest – was announced, it caused a stir within the artistic community. Some creators even saw it as a sign that "we’re watching the death of artistry unfold right before our eyes," as one Twitter user suggested. Others fear that the rise of generative artificial intelligence programs will contribute to the 'uberization' of creative industries.

Fears that were stoked by the San Francisco Ballet’s (SFB) recent campaign for "The Nutcracker." The American dance company used illustrations generated by Midjourney to promote the famous ballet... and many internet users were not impressed.

Several turned to social networks to criticize the SFB’s decision to use an AI image generator instead of a "flesh and blood" illustrator. "So disappointing to see that an organization in the arts is using AI art and actively harming visual artists," wrote one user under an SFB Instagram post. "AI art in its current form uses work from artists without their consent, credit, or compensation."

Confronted with this rising interest in AI image-generating programs, some artists are mounting a resistance. One of the places where they are expressing their dissatisfaction is on the ArtStation platform, where many of them host their work. It all started when some of them noticed in December that images created by artificial intelligence software were being uploaded to the platform.

Something that’s unacceptable in the view of Suzanne Helmigh, art director at Ghostfire Gaming. “I am really considering removing my art online on places such as Artstation and Deviant Art,” she said on Twitter. “If the point of creating art is lost and all our work is good for is to be fed into a machine to be abused and Frankensteined into some AI visuals.”

Images that protest and denounce the presence of these controversial works alongside artists’ works were uploaded to ArtStation with the hashtag #noaiart in support of the protests by digital artists who use the platform. ArtStation responded to the controversy with an FAQ about AI that includes the statement that “ArtStation’s content guidelines do not prohibit the use of AI in the process of artwork being posted.”

The Epic Games subsidiary added that it does not want “to become a gatekeeper with site terms that stifle AI research and commercialization,” while stressing that artists should be free to decide how their creations are used. Some image banks like Shutterstock have sided with ArtStation by welcoming AI-generated images with open arms, while others like Getty have banned them.

And what about the position of museums and other cultural institutions in all of this? Are they ready to host images generated by DALL-E 2, Midjourney and others, alongside works by Leonardo da Vinci, Pablo Picasso and Yayoi Kusama? The idea is still under debate since it involves ethical questions about the conception of these images, transparency regarding the process as well as the consent of the creators whose works are used to train the artificial intelligence software.

Additionally, questions arise regarding the machine’s role as an actor or as a passive technical element: is it in control of the creation or just another tool for visual artists, not unlike the camera?

The Museum of Modern Art in New York sees this technology offering an opportunity to showcase its own rich collection in a new way. It has invited Turkish-American digital artist Refik Anadol to use the 200,000 or so works of art in the museum’s possession to illustrate the history of Modern art, with the help of artificial intelligence.

Visitors to the New York museum can view the final result, until March 5, through several immersive installations, collectively called "Refik Anadol: Unsupervised." These installations take the form of three protean data sculptures that show the "dreams" and "hallucinations" of the machine, as Refik Anadol explains.

Ever-changing dreams that mirror art, according to Michelle Kuo, curator at MoMa and co-organizer of the exhibition "Refik Anadol: Unsupervised." "This project reshapes the relationship between the physical and the virtual, the real and the unreal," she said in a statement.

"Often, AI is used to classify, process and generate realistic representations of the world. Anadol's work, by contrast, is visionary: it explores dreams, hallucination, and irrationality, posing an alternative understanding of modern art – and of artmaking itself.”

But, while they may be aesthetically arresting, the 'dreams' of the machine, of AI, are still nourished by those of real human creators. If artificial intelligence is pushing the limits of creative human expression further than ever, it still requires commands or prompts from a real person in order to create.