Every recommendation engine claims to know what you want before you do. The promise is convenience; the mechanism is a slow rewrite of what counts as good, interesting, or sincere.
This piece follows how AI companies repackage taste — mood boards, playlists, font pairings, color grades — into proprietary signals they can scale. When "looks human" becomes a product feature, designers and artists are asked to optimize for a detector, not an audience.
We interview practitioners who've watched briefs change from cultural references to prompt keywords, and researchers mapping how style classifiers reinforce the mean. The question isn't whether machines have taste. It's who profits when yours is inferred from theirs.
Placeholder section: additional reporting on training data labor, platform guidelines, and the political economy of "authentic" aesthetics in generative workflows will appear in the published version.