Machines do not plagiarize, people do

#draft

Machines do not get inspired or

Machines do not plagurize

They are not inspired, and they do not make references. I’m not saying this because they don’t reach an arbitrary aesthetic or experiential threshold of any of these terms. I’m saying it because it’s a fact based on the way these systems are built and trained, one that is ignored by arguments that equate an image generator [spawning] a new image from its data set with a human making art that was inspired, consciously or not, by art they’ve seen over their lifetime.

Consider the recent controversy when the Getty logo began appearing like a ghost of IP in the corners of some Stable Diffusion results. Reactions to this, from all sides, are framed by a pretty sophisticated, almost unconscious distinction between collage, reference, and authorial intention. For these images to be at all remarkable, they must be considered as fundamentally different than an image made by a human artist with a Getty logo in it. In a painting or a collage, it would be a reference or a detournment. In a body of student work or on a website, it would be sloppy plagurism. Rephotographed, ala [Sherri Levine], and exhibited, it would be cultural commentary and knowing appropriation. Coming out of a machine, however, we see it to be a crack in the facade, a mistake, and possibly proof of illegal use of copyrighted images.

It’s our interpretive stance that reads the machine’s role in this process as plagiarism (or inspiration). The machine left it in for the same reason machines considered [rulers to be a sign of cancer]: because it’s part of the right answer to the question of “what pixels are usually in pictures that have labels that match the prompt?”

If you made a giant pile of other people’s drawings, and scooped a bunch of them up, why would we assign the agency of plagiarism to the pile, and not the person who made the pile, or who did the scooping? If you built an automated scoop, that knew where some types of images are better than humans do, and asked it to retrieve something very much like a copyrighted work, still, why think the scoop is the plagiarizer rather than the tool of a plagiarizer?

This is not the only way we could be using this technology.

Of course, the massive elephant in the room is that we don’t have to use them this way [ link to the GTA essay]. Specific products, such at Midjourney or Chat-GPT are trained on the mountain of other people’s artwork and released to the public, but the underlying technology doesn’t require that we do that. Many artists, such as Holly Herndon and [[ Sebastian ]] use LLMs trained on data sets they themselves have created or compiled. Of course, that’s not always as fun or easy as a model that can work with general concepts, which is what most people mean when they talk about “the AI.” After all, building something like that is something an [art student] would do. But if we’re even a little pedantic about the distinction, we can see holes in a lot of the prevailing arguments and positions about the technologies. When Sam Altman or other directors of generative products talk about either the inevitability of the technology or safeguards they assure us they are taking, they are talking about their products, which were deliberately trained on certain problematic data sets. When they talk about the importance of slowing down AI development, they’re talking about putting seatbelts on the race cars they built when they could instead simply be selling kits to make your own horse and buggy.