What Are LLMs Doing to Scientific Communication? Measuring Changes in Writing Practices and Reading Experience

The study's most underappreciated implication is directional: the drift isn't random noise but a systematic narrowing, meaning models trained on future scientific corpora will inherit a compressed version of domain vocabulary and syntactic variety rather than the full distribution that existed before LLM-assisted writing became common.

This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It belongs, however, to a broader conversation about training data provenance and corpus quality that has surfaced repeatedly in debates around synthetic data use. The concern here is a slower-moving version of the same problem: not deliberate synthetic injection, but organic stylistic convergence that quietly degrades the signal diversity in scientific text. For anyone building retrieval systems or domain-specific models on top of arXiv-style literature, that distinction matters less than the outcome, which is a corpus that looks real but behaves differently than historical baselines.

Watch whether a major NLP benchmark maintainer, such as the ACL Anthology team, begins versioning or date-stratifying training splits to isolate pre- and post-LLM-adoption text. If that practice emerges within the next 12 months, it will confirm the field has accepted corpus drift as a reproducibility problem worth controlling for.