culture

AI writing assistants are producing a weighted average of existing prose — and as more writers use them as drafting tools, that average is becoming the baseline. What was once a diverse ecosystem of registers, cadences, and argumentative personalities is quietly narrowing toward a single smooth voice: competent, correct, and indistinct. The signal is now measurable. Phrases statistically overrepresented in LLM-generated text — "delve into," "it's important to note," "multifaceted" — appear in indexed web content at two to four times their pre-2022 frequency. Stylometric entropy, a measure of vocabulary and syntactic diversity across a corpus, has declined in several monitored publishing verticals. The homogenisation is not just a qualitative impression; it is showing up in the data. But the framing of "voice" may be imprecise. If human writers are themselves trained on a corpus — everything they have ever read — then the difference between a person developing a voice and a model doing so is more a question of mechanism and sample size than fundamental kind. The more precise loss may not be voice at all, but surprise: the sense that an argument could arrive somewhere the reader has not already been. A model optimising on existing text cannot, by construction, exceed the distribution it was trained on. Human writers, drawing on embodied experience that falls outside any corpus, occasionally can.