Signal Study Finds Rising LLM Stylistic Trace (Em-Dash Usage) in US Congressional Press Releases

Summary

A preprint by Przemysław Czuma investigates whether large language models leave measurable stylistic traces in text written with their help. It focuses on the most widely discussed trace, the em-dash, particularly its unspaced form. That form is common in typeset English prose but unusual in US press writing, where AP style calls for dashes with spaces on either side. Using a preregistered design, the study tests whether this trace is measurable in congressional press releases. It measures how the population-level frequency of em-dash use changed between 2021 and 2025, a period spanning before and after LLM tools became widely available. Rather than analyzing content directly, the methodology tracks a stylistic fingerprint to offer a quantitative view of how far AI-assisted writing has penetrated official government communications.