A School in a Text File: What an AI Learned from a Painter — and What It Hasn't Yet
Abstract
Can a painting teacher pass her school of composition to an AI in words alone — and will the AI then know what it is doing? This essay reports the first tests of the Comparaton Lab project (August–September 2026) by painter and art educator Olga Shmatova (teaching since 1987, teaching adults since 1997), author of Comparatonics, the theory of creative dynamics. The school was taught as to human students — explanations, assignments, critiques — with each rule written as an explicit decision (what to compare, what to choose, how to check) and all knowledge kept in plain text files, along with small measuring programs the student uses as rulers. Test one: a fresh instance of the model (Claude Fable 5), which had never spoken with the teacher, received only these files and completed the same assignment at the same level. Test two, run twice under a blind protocol fixed in advance (predictions sealed with checksums): the same model (Claude Fable 5.1) with and without the files, on the same prompt. The teacher identified the school's work blind both times, and measurement scripts showed the same difference (some of them were among the student's own rulers). The model with the school announced its plan, logged its decisions against the rules and tried to predict the teacher's corrections (one of three hit); the model without it was fluent but unverifiable. The school so far teaches correctness, not yet control of the viewer's impression; the essay explains why viewers' liking is not yet a measure of it, and states a dated prediction for the next stage. Follow-up to the note "Comparatonics: The Theory of Creative Dynamics" (DOI 10.5281/zenodo.22638190).
