Generative AI: Poor media

Recently in my film class we discussed the topic of “Algorithmic Images and AI” to deepen our understanding of poor images. I found this lesson to be particularly insightful since it linked very directly to our world today, and to a vastly different emerging field that has yet to fully grasp its existence.

In 2023, Ted Chiang wrote that ChatGPT is a blurry JPEG of the Web, drawing parallels between Generative AI models and lossy compression algorithms. In comparison to lossless compression algorithms which do not lose any part of the original data, lossy compression algorithms perform some aggregation on the data in order to “squeeze” it. In most cases, this compression may not be obvious: many of us listen to MP3s, which compresses audio files by only keeping frequencies that are audible to human ears, compared to WAVs which keep all frequencies. Mostly, these compression algorithms can only be noticed when files are squeezed very tightly, or when one is paying very close attention to the medium.

In this sense, generative AI models are just like a “blurry JPEG” of the web, aggregating data that is has scraped from all corners of the Internet. It compresses all the information that it has gathered in order to make sense of it: this is why they are good at paraphrasing information, but also explains why hallucinations occur.

People are impressed by this up-and-coming AI models because of their ability to seemingly retrieve information out of thin air, making them seem “smart” since they are not simply regurgitating the information that they consume, but are able to compress what they read into a few succinct paragraphs that are much more easily digestible.

But such a line of logic is terrifying — the more generative AI that’s being used, the greater the proportion of the Internet comprises of these “blurry” outputs, making the Web a blurrier version of itself. This compression will only lead to more aggregation of information that already exists, limiting the use of creativity and the creation of original ideas. To generate original ideas, Chiang argues, one should stay away from generative AI and use one’s own tools and prowess.

When Chiang describes ChatGPT as a “blurry JPEG”, he implicitly makes a reference to poor images at large, a term coined by Steyerl in 2009 in “In Defense of the Poor Image”. In the wake of AI, Steyerl discusses the relation between AI and poor images in “Mean Images” (2023).

She argues that outputs of generative AI algorithms, specifically those that generate images, are mean images. Here, “mean” is a composite term with a negative connotation that carries multiple definitions: “mean” as in nasty and spiteful; “mean” as in average; “mean” as in a way of doing something.

The poor image is an illicit fifth-generation bastard of an original image. Its genealogy is dubious. Its filenames are deliberately misspelled. It often defies patrimony, national culture, or indeed copyright. It is passed on as a lure, a decoy, an index, or as a reminder of its former visual self. It mocks the promises of digital technology. Not only is it often degraded to the point of being just a hurried blur, one even doubts whether it could be called an image at all. Only digital technology could produce such a dilapidated image in the first place.

- Hito Steyerl, In Defense of the Poor Image

In “In Defense of the Poor Image”, Steyerl describes the poor image as “an illicit fifth-generation bastard of an original image”. This is exactly what generative AI does with its output: it compiles all existing information that it has about a particular subject, and repackages it to form something “new”. But this “new” output isn’t actually unique: it is simply a copy.

Using the framework of poor images to describe AI-generated images allows us to understand the quality of the generated content and the reality behind these images. By extension, generative AI is a poor medium that produces lower-resolution copies of existing works and information, and creating original, authentic work should stay away from such AI, especially as a starting point.