“Woman reading a book under the night sky, sleepy atmosphere” – I type in the Text 2 Dream function of the Deep Dream dream generator. After less than a minute, an image of what I described comes back to me. Welcome to the world of AI image generation, where you can create what looks like top-notch artwork with just a few text prompts, even if in reality your skills are no more than drawing miniatures.
AI image generation seems to be everywhere: on TikTok, the popular Manga AI filter shows what you look like in Japanese comic style, while crowds of people use it to create images for everything from company logos to picture books. It had already been used by one major publisher: sci-fi publisher Tor discovered that the cover it had created used a licensed image created by AI, but decided to do so anyway “due to production constraints.”
Among the biggest players in the AI industry are companies such as MidJourney, Stable Diffusion and Deep Dream Generator (DDG). Up to a certain point, they can be used for free, which makes them attractive to those who want to try them out. There’s no denying that they’re fun, but a closer look at the images they produce reveals the quirks. The face of the woman in my painting has very strange features and seems to hold many books. The paintings also have a similarly refined, somewhat kitschy aesthetic. And while there is an initial thrill when the image comes along, there is no creative satisfaction.
The implications of AI image generation are far-reaching and could affect everything from movies to graphic novels and beyond. Child illustrators were quick to raise concerns about technology on social media. Among them is author and illustrator Rob Biddulph, who says that AI-generated art “is the exact opposite of what I think art is. Basically, I’ve always felt that the art is in translating something you feel inside into something that exists outside. Regardless of the form, be it a sculpture, a piece of music, a literary work, a performance or a painting, true art lies far more in the creative process than in the final work. And simply pressing a button to generate an image is not a creative process.”
There are deeper problems beyond creativity. As part of the #NotoAIArt online campaign, artists shared their concerns about the legality of AI image generators and how they could devalue the art of illustration. To create images from prompts, AI generators rely on databases with pre-existing graphics and text. They include billions of images that have been scraped off the internet. Among the largest is the open-source LAION-5B dataset used by DDG’s Text 2 Dream. Kaloyan Chernev, founder of DDG, says the dataset includes “largely public domain images from the Internet,” but many artists and illustrators say the databases often contain many copyrighted images as well.
Harry Woodgate, author and illustrator of Grandad’s Camper, which won the Waterstones 2022 Picture Book Award, says: “These programs rely entirely on the pirated intellectual property of countless working artists, photographers, illustrators and other rights holders.” This is a point echoed by illustrator Anoosh Syed: “AI doesn’t look at art and create its own. It samples everyone and then mixes them with something else.
While prompts for image generators can be very general, they can also ask for an image based on another artist’s work, further blurring ethical boundaries. Syed says this can lead to the creation of images that are “deliberately meant to imitate my style” or the style of other artists without their permission. There’s an argument that AI generators work much like humans when it comes to the impact of other people’s work, but Biddulph says, “The human artist also adds emotion and nuance to the mix, and memory – particularly its downsides.”
He adds, “If I’m creating an image and I decide it should be Hockney style, I’m not going to search the internet for millions of Hockney style images, figure out exactly what characteristics make those images Hockney style, and then apply them to my photos, systematically and with forensic accuracy. I’ll think, “I like the way Hockney has juxtaposed blocks of purple, green, and ocher in this painting of a field I saw in the National Gallery.” And then I’ll try to add it to my photo. I will inevitably remember it badly and will probably end up creating something that is a bit like something Hockney once painted, but in my own style.”
AI doesn’t look at art and create its own. It samples them all – then mixes them with something else
Syed agrees, saying “another person will never look at a painting exactly the way the original artist did. They will never move their hands like the original artist did. AI doesn’t do the same – it can only copy.” When a human artist ’emulates a style or presents a work of art as his own, this is incredibly welcome – and in some cases can be seen as copyright infringement. This is basically what AI does.”
Chernev says he is aware of the “complex ethical issues surrounding the use of non-public domain images and the potential impact on artists whose work is used to train AI tools like ours.” However, there is a more insidious danger: the possibility of creating potentially illegal images. Chernev admits that when Text 2 Dream was first launched, people tried to “generate images of naked children even though there were no such images in the training dataset.”
It adds: “As AI continues to evolve, there is a risk that it will be able to synthesize images of inappropriate or illegal topics from existing content. In response, we quickly adjusted our tools to prevent the generation of any inappropriate or illegal content, including nude images of children and NSFW material. We are committed to ensuring responsible and ethical use of our image-generating service.”
Although Chernev claims that DDG reported the incidents to the authorities because the whole AI image generation is unregulated, which artists are quick to point out. Both Woodgate and Dapo Adeola, who won Illustrator of the Year at the 2022 UK Book Awards, would like more regulation. “A nice first step,” says Woodgate, “would be to reject the UK government’s proposed copyright exception allowing text and data mining for any commercial purpose, and instead opt-in models.” In this way, he says, any future databases would be created using voluntary contributions that are paid accordingly.
Adeola agrees, saying that “the easiest thing is to get artists permission to use their work” along with a fee. Czernew says DDG accepts requests from artists who want to be excluded from their system, but the “ask for forgiveness, not permission” model doesn’t sit well with Adeola, who says seeking permission “should be the first step.”
While, according to the artists, children’s book illustrations will remain largely unaffected, AI-generated imagery could potentially eliminate the smaller works that aspiring artists often rely on to build portfolios. Syed says for things like fanart, self-published books, logos, and family portraits, people can turn to AI. “These customers tend to be more concerned with saving money than with the quality of the finished product,” he says. “They will prefer to use AI if it means keeping costs down. So many of these small jobs will disappear.”
The growing use of AI, says Adeola, will also lead to the devaluation of artists’ work. “For me,” he says, “there is already a negative attitude towards the creative industry. Something like this reinforces the argument that what we’re doing is easy and we shouldn’t be able to make the money we manage.” Biddulph goes further. “There is no question that AI-generated art is devaluing illustration,” he says. “People will obviously start to think that their ‘work’ is just as important as the work created by someone who has spent their entire career making art. This is nonsense, of course. I can use my iPhone to take a nice picture of my daughters, but I’m not Irving Penn.
At the moment, AI image generation is largely used for fun, but Chernev says it’s “fast approaching the level of sophistication and complexity that will allow it to generate very realistic and nuanced images. I am convinced that content generated by artificial intelligence can not only improve the work of artists and designers, but also enable the creation of completely new forms of art and expression.
Artists and illustrators are not so sure. “AI-generated art has a specific ‘look’,” says Syed. “As time goes on, users will become more used to it and start turning away from it due to its inauthenticity and ‘cheapness’. I also think we may even see the re-emergence and recognition of traditional media in response to AI.”
Moreover, illustrators firmly believe that their most honest critics and biggest fans – children and teenagers – will not be convinced by the art of AI. “Children’s books are very complex, multimodal forms of communication,” says Woodgate. “Kids who read them expect a lot, not only from the stories and illustrations, but also from the people who create them.”