How DeviantArt navigates the minefield of AI art

Artificial intelligence learns how to make art, and no one knows exactly how to handle it – not even DeviantArt, one of the most well-known artists’ houses on the internet. Last week, DeviantArt decided to enter the minefield of AI image generation, launching a tool called DreamUp that allows anyone to create images of text prompts. It’s part of a larger effort by DeviantArt to give human artists more control, but it’s also caused confusion — and anger among some users.

DreamUp is based on Stable Diffusion, the open-source image spawning program created by Stability AI. Anyone can sign up to DeviantArt and get five prompts for free, and people can buy between 50 and 300 a month with the site’s Core plans, plus more for a fee per prompt. Unlike other generators, DreamUp has one distinct quirk: it’s built to detect when you’re trying to mimic another artist’s style. And if the artist objects, that should hold you back.

“AI is not something that can be avoided. The technology is only getting stronger day by day,” said Liat Karpel Gurwicz, CMO of DeviantArt. “But having said that, we think we need to make sure that people are transparent about what they’re doing, that they respect creators, that they respect creators’ work and their wishes around their work.”

Unlike some newstell Gurwicz and DeviantArt CEO Moti Levy The edge that DeviantArt is not doing (or planning) any DeviantArt-specific training for DreamUp. The tool is vanilla Stable Diffusion, trained on all the data Stability AI had collected at the time DeviantArt adopted it. If your art was used to train the model DreamUp uses, DeviantArt can’t remove it from the stability dataset and retrain the algorithm. Instead, DeviantArt targets copycats from a different angle: prohibiting the use of certain artists’ names (as well as the names of their aliases or individual creations) in prompts. Artists can to fill out a form to request this opt-out, and they will be manually approved.

Controversially, Stable Diffusion was trained on a huge collection of web images, and the vast majority of creators disagreed with inclusion. One result is that you can often reproduce an artist’s style by adding a phrase like “in the style of” at the end of the prompt. It has become a problem for some contemporary artists and illustrators who don’t want automated tools to copy their signature look – for personal or professional reasons.

These problems also occur on other AI art platforms. Questions about permission have led, among other things, to web platforms, including ArtStation and Fur Affinity, to completely ban AI-generated work. (The stock photo platform Getty also banned AI art, but it’s collaborating with Israeli company Bria on AI-powered editing tools at the same time, marking some sort of compromise on the issue.)

DeviantArt has no such plans. “We have always embraced all types of creativity and makers. We don’t think we should censor any form of art,” says Gurwicz.

Instead, DreamUp is an attempt to mitigate the problems – mainly by limiting direct, intentional, unauthorized copying. “I think today, unfortunately, there are no models or datasets that have not been trained without permission from the creators,” says Gurwicz. (That certainly applies to Stable Diffusion, and probably also to other large models such as DALL-E, although sometimes the full dataset of these models is not known at all.)

“We knew that whatever model we were going to work with, this baggage was going to come,” he continued. “The only thing we can do with DreamUp is prevent people from also taking advantage of the fact that it was trained without permission from the creators.”

Like an artist is DeviantArt is fine with being copied, but will urge users to credit them. When you post a DreamUp image through DeviantArt’s site, the interface asks if you’re working in the style of a specific artist and asks for a name (or several names) if so. Acknowledgment is required and if someone flags a DreamUp work as incorrectly tagged, DeviantArt can see what prompt the creator used and make a judgment. Works that omit credit, or works that deliberately evade a filter using tactics such as misspelling a name, may be removed.

This approach seems usefully pragmatic in some ways. While it doesn’t address the abstract problem of artists’ work being used to train a system, it blocks out the most obvious problem that causes the problem.

Yet there are several practical shortcomings. Artists should know about DreamUp and understand that they can request to have their name blocked. The system is primarily focused on granting control to artists on the platform rather than non-DeviantArt artists who vocally object to AI art. (I was able to create works in the style of Greg Rutkowski, who has publicly expressed his disgust of use in prompts.) And perhaps most importantly, the blocking only works on DeviantArt’s own generator. You can easily switch to another Stable Diffusion implementation and upload your work to the platform.

In addition to DreamUp, DeviantArt has rolled out a separate tool intended to address the underlying training question. The platform has added an optional flag that artists can check to indicate whether they want to be included in AI training datasets. The ‘noai’ flag is intended to create certainty in the dark scraping landscape, where artists’ work is usually treated as fair game. Since the design of the tool is open-source, other art platforms are free to adopt it.

DeviantArt does not do any training itself, as mentioned earlier. But other companies and organizations must respect this flag in order to comply with DeviantArt’s terms of service – at least on paper. In practice, however, it seems above all ambitious. “The artist will very clearly signal to those datasets and to those platforms whether they’ve given their consent or not,” says Levy. “Now it’s up to those companies whether they want to put in the effort to search for that content or not.” When I spoke to DeviantArt last week, no AI art generator had agreed to respect the flag going forward, let alone retroactively remove images based on it.

At launch, the flag did exactly what DeviantArt hoped to avoid: it made artists feel like their consent was being violated. It started as an opt-out system that allowed training by default and asked them to put the flag if they objected. The decision probably didn’t have much immediate effect as companies scrapping these images were already the status quo. But it enraged some users. A popular tweet by artist Ian Fay called the move “extremely scummy”. Artist Megan Rose Ruiz released a series of videos criticism of the decision. “This is going to be a huge problem that will affect all artists,” she said.

The outrage was particularly pronounced because DeviantArt has offered tools that protect artists from another technology that many are ambivalent about, namely non-fungible tokens, or NFTs. A program was launched last year and has since expanded to detect and remove art that has been used for NFTs without permission.

DeviantArt has since then tried to tackle criticism of its new AI tools. It has the “noai” flag enabled by default, so artists must explicitly state that they agree to have images scraped. It has also updated its terms of service to explicitly order third-party services to respect artists’ flags.

But the real problem is that, especially without extensive AI expertise, smaller platforms can only do so much. There are no clear legal guidelines for creators’ rights (or copyright in general) for generative art. The agenda has so far been set by fast-moving AI startups like OpenAI and Stability, as well as tech giants like Google. Besides simply banning AI-generated work, there’s no easy way to navigate the system without touching what has become a third track for many artists. “This isn’t something DeviantArt can solve alone,” admits Gurwicz. “Until there is the right regulation, these AI models and platforms need to go beyond what is required by law and think ethically about what is right and fair.”

For now, DeviantArt is making an effort to push that mindset – but it’s still working out some major kinks.