I can't tell who's teaching who anymore

This essay examines how language circulates between people and large language models, arguing that feedback loops have blurred authorship and eroded linguistic specificity. It recounts research showing human reviewers nudged models toward particular words (Florida State’s finding about the jump in use of “delve”) and a Max Planck analysis of 737,000 hours of podcasts linking ChatGPT’s release to the spread of model-favored phrasing. Repeated training on model output produces “model collapse,” averaging away rare, particular turns of phrase, while detectors flag clean, evenly paced sentences and certain buzzwords. The writer keeps a “Literary Graveyard” of em dashes and favored words she now strips because they read as machine-generated.

On a personal level, the piece describes stripping the clear, structured style taught by the writer’s mother to avoid being misidentified, producing self-surveillance and an identity crisis. The argument distinguishes careful, human-guided drafting (including model-assisted revision) from unedited model dumps and insists value lies in the effort spent. Concrete consequences surface: documentation traffic dropped and engineers were laid off as tools absorbed knowledge. No easy fix is proposed; the conclusion is an unresolved anxiety about whether the slow skills that defined a career will be smoothed away.