The June 1893 issue of The Studio, a London arts newspaper, was devoted to a pressing question: “Is the Camera a Friend or Foe of Art?” Does the fact that a computer can produce a beautiful, lifelike image at the touch of a button diminish the value of “real” art? It’s the same question that artists are posing about artificial intelligence 130 years later.
AI has transformed every sector of the arts in ways so rapid and nuanced, varied and vast, that keeping track of it all has been nearly difficult. The influence of AI, with its quickly expanding ability to generate images, sounds, and words, is already obvious as the defining change in this decade’s cultural life.
What is less certain is whether this transition should be celebrated or feared.
The visual arts
A photo that wasn’t a photo won the Creative category of the Sony World Photography Awards last month. As lovely as it is spooky, Pseudomnesia: The Electrician is a sepia image of two unreal women. Boris Eldagsen, a 53-year-old German photographer, created it with DALL-E 2, a year-old AI program that makes images in response to spoken instructions; a word paints a thousand pictures.
Eldagsen entered the contest “as a cheeky monkey” with his image. He breached no rules in doing so, if only because no one had thought to update them for the AI age. “I’m expecting photo competitions to do their homework,” he says. When he found out he’d won, he wrote back to decline the award, describing how he’d created his “photo.” Sony nevertheless declared him the winner.
So he rented a tux and traveled to London for the awards event, where he intended to raise awareness about AI art. (It is art, he claims, but in a new form that does not belong in photography competitions.) When he realized he wouldn’t be able to give a speech, he climbed onstage and grabbed the microphone to explain why he was turning down his prize. The show continued as usual after an unpleasant moment.
Following a burst of media attention that spurred a controversy regarding AI art, the awards’ organizers issued a statement alluding to “creators engaging with lens-based” work. It avoided the most important point: Eldagsen had made his image without using a lens at all.
Making AI graphics from word prompts, or “promptography,” as Eldagsen refers to it, takes skill. It’s like photography: anyone could press the shutter button on a camera at random and produce a stunning image. A true artist can do it again and over again. Eldagsen’s technique consists of 11 levels of unique text prompts that cover all aspects of a composition. He instructs classes in it.
Creating a terrific image still requires effort, but creating a bad one has never been simpler. There has always been a gap between trained and amateur efforts, but Eldagsen believes that with AI, “professionals are scared because that gap is becoming smaller.” “The crappy images of the past will fade away.”
“There’s an issue that’s more important in the background here,” writes Kate Crawford, author of Atlas of AI. When it comes to creating AI images, “the ‘creativity,’ if you will, is premised on a massive capture of the commons.” We lost our grazing land in the 18th century; we lost our data in the twenty-first.
AI image engines like Midjourney and DALL-E 2 are trained by “scraping” – an unappealing, but perhaps suitable, violent term – millions of images and captions from the internet. “Every time you upload a photo of a friend on vacation or engage in any kind of publicly visible online activity, that is now harvested into these massive datasets,” Crawford explains.
People ask, ‘Is this art?’ rather than examining the fundamental practice itself. Is this the kind of world we want to encourage, where everything is taken without consent?”
In January, cartoonist Sarah Andersen joined a class-action lawsuit against the London-based business behind Stable Diffusion, an image maker that began last August and may be instructed to create images in the manner of any artist, including Andersen. Getty Photographs filed its own lawsuit in February, claiming that Stable Diffusion created photographs that are nearly identical to its own copyrighted shots, with some even copying the watermarks that Getty uses to prevent infringement.
LAION-5B, a massive library of scraped photos (the “5B” stands for five billion), was used to train table Diffusion. In the United States, mass scraping without previously obtaining authorization has been ruled lawful as “fair use” thus far. Some people are unconcerned about this. “Personally, as an artist, I came to the conclusion that I can’t protect the images I’ve created in my life,” Eldagsen adds sarcastically. Others, however, are uneasy. Last year, two artists, Mat Dryhurst and Holly Herndon, launched haveibeentrained.com, which allows you to see if your images have been used in AI training sets. LAION-5B swiped an image taken by her doctor for her confidential medical files, according to one woman.
Since Stable Diffusion agreed to collaborate with haveibeentrained.com to allow consumers to “opt out,” over one billion images have been removed from its training set. Perhaps this is progress. However, the “grab everything first, then put some of it back if you ask nicely” strategy is not the only way to go. Adobe introduced a competing AI image-maker in March, trained only on non-copyrighted images and those to which Adobe already has the rights.
Even if the training data is obtained legally, many artists concerned about their jobs would prefer to boycott AI.
Much of the concern is centered on genres where an image’s value as art is secondary to its function for something else: to promote a product, communicate information, or brighten up a bit of prose – areas where “good enough” is more important than “good”.
Illustrators are especially concerned. “This is effectively the greatest art heist in history,” they warned in an open letter signed by hundreds earlier this month. “If this seems alarming, remember that AI-generated work has already been utilized for book covers and editorial graphics, replacing illustrators from their livelihood… Generative AI art is vampirical, feeding on previous generations of artwork while sucking the lifeblood of living creators.”
The Cinema
Text-to-video Last September, Meta introduced Make-A-Video, a digital application that can produce short silent movies from a single sentence. A week later, Google launched Imagen, a competitor product. Both are still only available to a small number of human testers, but Hugging Face, an open-access AI organization, made its own text-to-video tool, ModelScope, available to anyone in March. It quickly went viral, driven by the line “Will Smith eating spaghetti,” in which the actor looks to be gnawing fistfuls of pasta like a man possessed.
Luhrmann believes that AI helpers and avatars will become commonplace and that humans will cease cringing at them. It’s just like mobile phones, he says over a martini after a recent event at London’s Design Museum. “When they first arrived, they were only used by plumbers and tradespeople.” It was as though to say, “How gauche!” AI “will just feel human” in time.
But not overly human. Luhrmann had previously stated onstage concerning AI, “I’m not terrified as a creative person… If I told an AI, ‘Write me a screenplay in the style of yours truly about King Lear,’ it would lack compassion. The defects, the imperfections, are what make us human.”
Actors who are not covered by such a clause may be disturbed by the work of businesses like Flawless AI, whose AI-powered editing might effectively reduce them to ventriloquists’ dummies. It will be useful for rereleasing films in foreign markets. Has it been dubbed into another language? They can sync an actor’s lips to fresh dialogue with no issue. Is there foul language? Nothing could be simpler than substituting “fiddlesticks” for an F-word. Is it necessary to replace a line denouncing China with one supporting the CCP? In theory, they could do the same. (Flawless AI denied an interview request from The Telegraph.)
Visually modifying an actor’s mouth would be useless if AI couldn’t also change their voice. However, AI can produce uncannily convincing human noises. It speaks, sings, and performs symphonies.
