This month, the internet was flooded with stunningly ethereal digital art portraits, thanks to the work of the latest artificial intelligence-assisted application to go viral: Lensa. Users uploaded their photographs to the app and then – for a small fee – it used AI to transform their profile pictures into, say, a magical elfin warrior princess version of themselves, in no time at all.
This year has seen a breakthrough for AI-driven image generators, which are now better than ever in quality, speed and affordability. The AI models are “trained” on millions of pieces of image and text data scraped from publicly available content online, and as in the case of Microsoft-backed DALL-E, can turn short text prompts such as “Ronald McDonald performing open heart surgery” into unique images.
Anyone can now produce professional-looking images tailored to their desires, without having any training in art or design themselves. If that sounds great to you, you might not be one of the millions of humans whose livelihoods depend on being able to exchange those skills for money.
Those working in the more cognitive creative industries have long felt that they had nothing to fear from automation. After all, how could a computer ever recreate the aura of a masterpiece by Leonardo da Vinci, or possess the unique skill set required to devise a compelling visual marketing campaign for a luxury brand?
Early images generated with these tools were full of glitches that marked them out as machine-made. But as the results have become more convincing, creatives have grown more concerned. On the frontlines of this debate are gig workers such as graphic artists and commercial illustrators, who take art commissions based on client specifications.
Anyone inclined to dismiss the idea that AI could take over creative jobs as scaremongering should know: it is already happening. This winter, San Francisco Ballet used the independent research lab Midjourney to create the visual campaign for its production of The Nutcracker (although a representative for the ballet said that, despite using AI, nearly 30 human designers, producers, and creatives were also employed in the campaign’s making).
Another threat to artist livelihoods comes from these tools’ ability to create imagery “in the style of” specific artists. This functionality is fun when used to conjure up quirky visions of how Van Gogh might have painted Rishi Sunak riding into No 10 on a unicorn, but when it comes to living artists who have spent years developing their own distinctive style, the AI’s uncanny ability to mimic, without credit or compensation, becomes problematic.
Earlier this year, fantasy art illustrator Greg Rutkowski found out that his name was one of the most popular prompts on the AI platform Stable Diffusion – more popular than Picasso or Leonardo. “The only thing that could at least stop feeding the algorithm is to stop posting your work on the internet, which is impossible in our industry,” says Rutkowski.
The legal recourse for artists who feel these tools are infringing on their copyright is knotty and unclear. In the EU, lawyers are contesting the legality of using images under copyright for training AI models but as the UK bids to become an industry leader, it has already proposed a bill to allow carte blanche AI training for commercial purposes. Meanwhile it remains unclear if traditional copyright even applies here, as it is difficult to copyright a visual style.
While these issues have only recently garnered mainstream attention, there are factions of artists who predicted this when the field was still in its infancy, and have been working to develop solutions. Among them are Berlin-based artists Mat Dryhurst and Holly Herndon, who have created a search function that anyone can use to see whether their work has been scraped for a 150-terabyte dataset called LAION, which is used to train most AI image generators. Their organisation, Spawning, is also developing another tool that would allow artists to set permissions on how their style and likeness can be used by the algorithms, including the option to opt out entirely.
Both Stability AI – the organisation behind Stable Diffusion – and LAION have committed to partner with Spawning to honour consent requests made in advance of the next training of Stable Diffusion, and a recent update to the tool removed the ability to write prompts that specify an artist by name.
There are other flaws in the vast open datasets on which the AI models are trained, which limit its potential. Deficiencies in the diversity of the data, as well as biases held by the humans who originally labelled the images it learns from, have unwittingly coded the models with harmful stereotypes and representations. Some users are finding that Lensa creates overly sexualised female avatars, exaggerates racial phenotypes in its outputs, and has difficulty reading mixed-race features. Such issues might give pause to anyone thinking of using the technology for commercial purposes – at least until the training datasets are improved.
Many artists remain unfazed, and in fact believe the technology could open up possibilities for them to make better work, or at least to work more efficiently. Though she has not used it yet, the UK-based illustrator Michelle Thompson sees potential in the idea of using AI both to develop concepts and to refine artistic outputs. “I see it less as a threat and more of an opportunity,” she said, adding: “Like everything else, there will always be artists who can use the tools better.”
These tools are only as good as the datasets they are trained on. Human imagination, on the other hand, has no limit. For Dryhurst, AI models “could attempt to make a pale version of something we did years ago,” but that “doesn’t account for what we might do next”.
For those watching closely, the visual outputs of these widely available AI tools are already getting repetitive, and even untrained eyes will learn soon to recognise the hand of the machine. Some of the most interesting and conceptually rich work being made with AI is still coming from artists such as Mario Klingemann and Anna Ridler, who are customising their own training datasets, and curating the machine outputs in imaginative ways.
The kind of artificial intelligence we might imagine replacing artists – an entirely autonomous creative robot capable of human-like imagination and expression – does not yet exist, but it is coming. And as AI becomes more ubiquitous, artists, illustrators and designers will ultimately be set apart not by if, but by how, they use the technology.