Why AI won't stop saying 'shimmer'—and what it reveals

The article traces a pattern in how AI language models produce their own linguistic habits—overused words that emerge from the mathematical structure of the models themselves. Jill Walker Rettberg's observation that words like "shimmer" occupy unexpectedly dense points in latent space (the abstract vector space where language models store learned meanings) explains why they appear so frequently: they are, in a sense, mathematical hubs with more pathways leading to and from them than ordinary words. Their semantic ambiguity makes them especially useful bridges between concepts, turning them into convenient filler when models need to connect disparate ideas.

This quirk has a visual parallel in Google's design choices. The company's near-universal adoption of the sparkle icon—nearly 100 instances by 2024—was validated by a UX study showing that users recognized the icon as a sign of AI involvement, even if they could not articulate what specific action or capability lay behind it. Both the word "shimmer" and the sparkle icon communicate "something special is happening here" without committing to a concrete meaning, a quality that actually mirrors the opacity of large language models themselves. The article suggests this vagueness is not accidental but structural: black-box AI systems produce signals—linguistic and visual—that are deliberately or functionally ambiguous because the systems themselves do not always have a crisp explanation for their outputs.