One thing that isn’t well understood about the backlash against AI is that this new technology can feel intrinsically unfair to the people who made it possible. I’ll demonstrate this with a somewhat complex example.
For decades now, independent and professional artists have been sharing their work online, using sites like DeviantArt, ArtStation, Pixiv, or Twitter. When sharing their work, they assume — naturally — that the viewer will understand that this work is being shared to enjoy, not shared to be repurposed or used commercially. In many cases the websites being used to share this art explicitly point out whether or not the viewer has the rights to repurpose, re-upload, or otherwise exploit somebody’s art.
Anyway. When a great piece of art gets shared on social media, it spreads virally. Sometimes the artist themselves kickstarts this with careful use of hashtags, other times a popular user will retweet a work or link to it, bringing it to a wider audience. Through this virality some users will discover ‘hey, I haven’t seen this artist’s work before, and I really like it’, and decide to follow them on social media, or even buy merch off of their shop.
But let’s say you’re looking for something specific. You’ve just gotten into a new TV show or video game or comic book or even a series of novels. You want to hunt down new fan works around that franchise, or works around a specific character you like.
For prose, there are sites like Archive of Our Own, where people openly post fan works and tag them to make it easy to discover them, but the same is not true for art — illustrations are often posted with few or even zero tags to communicate what’s in the illustration, which can make it hard to discover new fan works for something you’re just getting into — you might need to know specific fandom shorthand or foreign languages to find what you’re looking for. This is partly a consequence of services like Twitter limiting your word count, but also, sometimes artists just don’t tag their work.
The online community came up with a workaround for this, imageboard websites like Danbooru. Imageboard users go through the feeds of various artists — not just popular ones, but often smaller up-and-coming artists and complete amateurs — and catalog their work, carefully uploading it with detailed tags and sourcing information. As you might imagine, this is a thrilling piece of technology to use, you can just type in your favorite character or franchise and see a near limitless feed of art by people from all over the world. And if you find an illustration you like, a link to their twitter or patreon (where you can repost it or support them) is right there. These sites even offer integrated support for translations, so if an illustration has some Japanese or Korean or Chinese text in it, you can mouse over it and see a translation that another user attached to it and understand the work. Seems great! But it’s not so popular with the artists themselves. Why is that?
This comes back to the assumption I mentioned earlier. When sharing their creative work, artists assume that the viewer will basically follow the implicit rules — look, don’t touch. I’m sharing this for you to enjoy it, not Do Stuff with it. You shouldn’t be putting this on a t-shirt without my permission, you shouldn’t be selling prints of it, you shouldn’t be uploading it on your website without my permission. These implicit rules are pretty logical if you understand how copyright works.
The imageboard catalogers are busy doing something they might view as a public service, while at the same time actively disregarding the desires of the artists who make the imageboards possible. At the same time, the imageboard potentially sends new viewers to the artists, helping them get discovered. Reconciling this might be difficult, but at least in this case, you can choose to believe that everyone involved wants artists to thrive, and wants to share and discover great art.
Anyhow, one of the for-profit AI services out there is called “Novel AI”. I bring this service up in particular because: Instead of writing freeform text prompts to generate an image, you give it a selection of precise tags to describe the image you want. It understands a very, very broad set of tags, and even understands how to imitate the styles of particular artists quite well — not just famous artists, but amateurs with online followings too. Not only can it imitate famous artists, but it can recreate almost any character you can think of from almost any franchise, even unpopular ones. That’s interesting. How did they do that?
All the volunteers helpfully copying artists’ work onto imageboards did that, unintentionally. They catalogued vast quantities of art, tagged it in detail, and classified it, while also re-uploading it onto servers with minimal access control. This made it straightforward for AI companies to come in and slurp up all this ready-for-training image and tag data and turn it into a model to resell. The tags used by this service are the exact same tags used by Danbooru, with some minimal effort taken to disguise it (for example, artist tags are hidden in suggestions, but work if you type them in.)
Suddenly the people who might have thought they were ‘just helping’ artists get discovered actually are responsible for helping replace those artists with for-profit AI. The imageboard catalogers weren’t getting into this to help make Novel AI’s creators rich, they probably just wanted to share and discover art. The artists who shared their work in the first place certainly weren’t getting into this to help make AI people rich, they wanted to create and share art.
But ultimately, some people broke the rules a little bit, maybe they thought they were doing a good thing… and then someone else came along and broke the rules a little bit more, and finally at the end of this chain of rule-bending we have a machine that gives you reasonable-fidelity knockoffs of specific artists’ styles and intellectual property at a price of cents-per-image, instead of the hundreds of dollars USD a commission from a working artist would normally cost.
You can see how it seems a little unfair, right?