What happens to ground truth when finding a factoid or photo no longer means consulting an archive but generating one from scratch? That’s the question that drives “Honey, AI Shrunk the Archive,” an essay I wrote for the forthcoming anthology New Directions in Digital Textual Studies.
A lot of ink has been spilled about whether generative AI is more like a sycophantic intern, a growing child, or a statistical songbird. But understanding how AI might transform the role of the archive, and in turn the production of knowledge, may require appealing to other metaphors. “Honey, I Shrunk the Archive” examines simpler, non-organic systems that share some mathematical features of large language models.
The essay argues that we can even learn something about generative AI by comparing it to something as simple as a cup of coffee. A coffee’s energy and temperature turn out to have mathematical correlates within large language models.
The most telling mechanistic analogy may be to a compression algorithm like JPEG or gzip. Similarly to the way JPEGs smooth over the fine details of a photo to save space, transformers distill vast archives into a compressed version of knowledge. In doing so, however, AI risks erasing the anomalies, the noise, and the outliers that make archives valuable sources of ground truth.
By shedding anthropomorphic metaphors and viewing AI as a statistical engine, we can understand why generated content is a threat to what makes archives irreplaceable: their ability to surprise, challenge, and tell the messy truth.
The forthcoming volume is edited by Christopher Ohge and Kristen Schuster and is scheduled for publication by Bloomsbury Press in 2025.