WHAT VOICE ACTUALLY IS
Linguists call it idiolect — the fingerprint of word choice, sentence rhythm, and syntactic habit unique to each writer. Forensic linguists have used idiolect to identify Unabomber Ted Kaczynski from his manifesto and to attribute disputed Federalist Papers to Madison versus Hamilton.
THE TRAINING-DATA REGRESSION
Large language models are trained to predict the most probable next token across billions of documents. Their output gravitates toward the statistical center of human prose — the median sentence, the unsurprising adjective. Students who edit with these models are pulled toward that same center.
THE FLYNN EFFECT, INVERTED
For most of the 20th century, measured cognitive scores rose generation-on-generation — the Flynn effect. Since roughly 2010, several countries have recorded reversals in verbal reasoning and reading. The mechanism is contested but screen-mediated language exposure is a leading suspect; LLM-mediated writing is a new vector on the same axis.
WHY SCHOOLS MISS IT
Plagiarism detectors compare text against a corpus of existing documents. AI-generated prose is novel by construction — it isn't copied from anywhere — so it passes. The detection problem and the homogenisation problem are different problems; tools built for the first are blind to the second.
THE OLDER PRECEDENT
Every literacy technology has reshaped cognition. Plato's Phaedrus warned that writing itself would weaken memory; the printing press standardized spelling and killed regional orthographies; spell-check eroded the felt difference between affect and effect. The pattern is offload, then atrophy — the question is what is being offloaded this time.
WHAT GETS LOST
Voice is not ornament — it is the trace of a thinker working through a problem in real time. When prose converges to a median, what disappears is not style but evidence of thought: the false start, the hedge, the unexpected metaphor that signals a mind reaching past its training.