Unauthoritative Pronouncements

John Siracusa has a good post on his blog summarizing some of the concerns regarding generative AI. There’s a lot of other kinds of “AI” in the news these days, but this is specifically about material wholly generated from text based prompts, using computer models trained on other images.

This question is at the emotional, ethical (and possibly legal) heart of the generative AI debate. I’m reminded of the well-known web comic in which one person hands something to another and says, “I made this.” The recipient accepts the item, saying “You made this?” The recipient then holds the item silently for a moment while the person who gave them the item departs. In the final frame of the comic, the recipient stands alone holding the item and says, “I made this.”

This comic resonates with people for many reasons. To me, the key is the second frame in which the recipient holds the item alone. It’s in that moment that possession of the item convinces the person that they own it. After all, they’re holding it. It’s theirs! And if they own it, and no one else is around, then they must have created it!

The act of creation is a tricky thing to pin down with people, because someone may not realize the ways in which they were influenced by something they saw before. It ought to, theoretically, be easier to pin down sources (the training data) a model uses, but the people making the popular models right now would really prefer if you didn’t.

If you feed a prompt into a model, you post the result as your own, and you get a cease and desist letter in the mail, then you suddenly flip-flop on who’s responsible for this image. Instead of you creating AI art with your carefully crafted prompt, the infringing work is the fault of an opaque data set you couldn’t possibly be held liable for.

It’s also not just copyright you have to worry about, but using bits and pieces of images from a dataset including CSAM, or other morally repugnant things you would not consciously include in your work. There are people who are looking at those things, for $2.20 an hour.

Do all the people that used Stable Diffusion during the time it had a tainted dataset need to post a disclaimer on the image they’re claiming ownership of? No, no one’s going to do that, because that’s not their fault.

The Gray Goo

The thing that really raises my alarms is how companies are training people to use generative AI to fill their social networks, publications, and sites with untraceable, generative AI images. In fact most of this post was originally part of a draft titled ‘Gray Goo’ I started in December when companies were bragging about their end of year AI progress, such as this boastful post from Meta.

In science fiction, the concept of gray goo exists to describe a hypothetical scenario in which self-replicating machinery outcompetes and replaces organic life. Generative AI models are not currently self-replicating, but we are feeding generative AI output into other algorithms, like image search results, or social network feeds.

If you look at the Explore tab on Instagram you have an algorithmic feed of thumbnails containing images. If you tap through into those you might notice that a few of the accounts posting the images are aggregators. They say things like “DM for credit or removal” as if the person who posted it didn’t know how they got their hands on the image in question.

These are usually topic based, like certain dog or cat breeds, desk setups, architecture, film photography, or whatever. The aggregators can also have links in their profile to shirts, or other merchandise or services, that they are selling to collect money for their hard work in reposting other people’s stuff.

Spend some time looking around and you’ll also see some Midjourney or Dall-E-looking images that may or may not have #AI on the post, or have already been laundered through one of those aggregators and contain no disclaimer or sourcing. (A word of warning, the #AI hashtag on Instagram is seemingly unmoderated and contains some anatomically questionable material. (No, I’m not just talking about fingers.))

The ultimate goal of these platforms is to have stuff, and not just any stuff, but filler between ads that they serve. People can see and mimic popular posts, and memes, on social platforms, but that still requires labor. Reduce the labor and you can increase the amount of new filler. That includes different variations on filler to fit new formats, like image tools to make horizontal aspect images vertical to increase Stories, or add animation to increase the amount of video filler, which means more video ads.

This also increases the number of people that wouldn’t feel guilty for uploading some untraceable generated stuff, because they are not knowingly copying anyone, like those aggregators are. They get the serotonin without the fuss, or the guilt.

So, it’s not just “I made this” of the user and their generated image, but “I made this” of the companies who want to sell ads against the endless stream of “user generated” images that they have safe harbor protections from. After all, if there is a problem with an image, there are existing moderation tools (that mostly don’t work, but whatever!) to deal with moderating things individual humans need to actively file objections to.

Consider why companies want you to use generative AI tools. It’s not altruism about democratizing expensive software.

Unique vs Ubiquitous

You can’t really put the genie back in the bottle here, but you can regulate the fuck out of that genie’s datasets, which will make it much less attractive to exploit people. Back to John:

In its current state, generative AI breaks the value chain between creators and consumers. We don’t have to reconnect it in exactly the same way it was connected before, but we also can’t just leave it dangling. The historical practice of conferring ownership based on the act of creation still seems sound, but that means we must be able to unambiguously identify that act.

There is also the possibility that people will lose interest in ubiquitous prompt-based tools because the work produced may be less special, or unique. That might also lead to a shift away from prompts that generate the whole enchilada to more of the tools that assist in editing and image manipulation that’s more human-centric.

Art is often a reaction to current trends in art. Everyone is painting realistic stuff? Let’s paint surrealistic stuff! Photography is clean, and crisp, and digital? Here’s the grainiest film you’ve ever seen in your life!

If the datasets can’t absorb current trends, or changing tastes, then they can’t easily be used for sole authorship of unique works that ultimately need to appeal to humans, not models.

Art will, uh, find a way.