Why Large Language Models Seem to Overuse Em-Dashes
Evidence, Mechanisms, and Limits of a Widely Noticed AI-Writing Signal
fun fact: this page contains zero em dashes.Abstract
The popular claim is that LLMs overuse em dashes. This review narrows it to a testable question: under what conditions do some outputs use the mark more often than a matched human baseline? The pattern shows up in some model outputs and in a large post ChatGPT scientific corpus, but rates vary by model, instruction and genre. The mark can be a salient cue. On its own it is not evidence of AI authorship, and no study yet establishes a universal cause.
A revision cue, not an accusation
People have used the em dash for centuries. House style, editors and software can all explain it, and corpus trends can't attribute a single document. The practical move is to treat a cluster of them as a reason to reread, not as forensic proof.
What's inside
- 01What has been measured. Corpus studies, cross model comparisons and a peer reviewed counterexample.
- 02A cautious mechanistic account. Autoregressive bridges, post training and the Markdown hypothesis.
- 03Why readers notice. A visually loud mark with a rhythm that can feel templated.
- 04A decisive next study. A preregistered factorial design across models, humans and genres.