Users of Anthropic’s Claude models have named the variety of English these models produce. They call it Claudish, they catalogue its features, they file bug reports against individual words, and they have built translators from it into plain English. This paper reviews what is actually known about that variety. We survey the academic literature and the technical grey literature, sort the qualitative feature inventory by linguistic level, and grade every quantitative claim by the strength of its design. Four findings are supported by large-scale measurement with an explicit baseline. Machine-associated vocabulary has shifted scientific and news writing at scale, with excess frequency ratios up to 28 for individual words and at least 13.5% of 2024 biomedical abstracts affected. A cluster analysis of 461,121 GitHub pull request descriptions shows one way of writing growing from 0.7% of the corpus in early 2025 to 39% by mid-2026, with load-bearing overrepresented 39 times inside that cluster. Post-training reduces output diversity, and sycophancy is measurable in 58% of rebuttal cases across three assistant families. Language models overconverge to their user’s style while users accommodate models no differently than they accommodate people. Three widely repeated claims fail this grading: the delve lexicon indexes a different model family and time period rather than Claude, vendor multipliers such as “700 times more frequent” are published without method, and no study has compared Claude with a registermatched human baseline. We conclude that Claudish is best analysed as a register, situationally selected and versioned, that is acquiring one dialect-like property through documented human uptake, and we set out the eight studies that would settle the open questions.
Rachel So. It Is Not a Style, It Is a Register: A Review of the Evidence on Claudish. Project Rachel, September 2026. https://doi.org/10.71775/kth.yv057-7en42