This work-in-progress takes as its provocation the ways that image-generating AI systems struggle with typographical forms. Systems like DALL-E 3 generate images of letters--shapes in text-appropriate contexts--but some of them are shaped a good bit less like letters than others. How should we understand these typographical hallucinations? Is there something that DALL-E 3 "knows" about typography that we don't? Today's AI operates on data at Internet scale, and I approach these questions in part by invoking the universe of printed matter as a nineteenth-century analog.

Lisa Gitelman is a media historian whose research concerns documents, data, and techniques of inscription. She is a professor of English and media studies at New York University and the author of Paper Knowledge: Toward a Media History of Documents  and Always, Already New: Media, History, and the Data of Culture, both recently published in Chinese translations.