A punctuation mark is now a population-level fingerprint of the AI writing era.
Researchers analyzed 69,632 first-version medRxiv preprints deposited between 2020 and 2025, tracking how often Discussion sections contained at least one em-dash. Before ChatGPT launched in late November 2022, 4.23% of those sections used one. By 2025, that figure hit 20.3%. The overall post-ChatGPT prevalence settled at 11.58% — nearly three times the pre-LLM baseline — with an odds ratio of 2.96. The study was pre-registered on the Open Science Framework before any confirmatory results were computed, which limits the usual concerns about p-hacking.
The em-dash has long been flagged informally as an LLM stylistic tell — models like GPT-4 and Claude lean on it in a way that strikes human readers as slightly off. This is the first large-scale attempt to measure whether that anecdote shows up in actual published science. It does, and the gradient matters: the rise was gradual through 2023, then accelerated sharply in 2024 and 2025, tracking the broader adoption curve of AI writing assistants rather than a one-time shock.
The authors are careful to note the study cannot establish causality — the em-dash is a population-level signal, not a per-paper detector, and other writing-habit shifts could theoretically explain part of the trend. Still, falsification tests against a placebo split in the pre-LLM era showed almost no change (+0.13 percentage points), and boilerplate sections of the same papers showed no comparable rise. If something other than LLM assistance is responsible, the study hasn't found it.