The Em-Dash as an AI Fingerprint? Congressional Press Releases Doubled Their Use in 2025
AI-assisted writing leaves a widely discussed “fingerprint”: the em-dash. A preregistered study analyzed more than 140,000 press releases published by U.S. congressional offices from 2021 to 2025 and found that the density of unspaced em-dashes rose sharply in 2025, reaching more than double the four-year baseline, providing quantitative evidence of large language model (LLM) penetration into official writing.
- The study analyzed 146,239 congressional press releases from 480 House and Senate offices (2021-2025)
- Unspaced em-dash density stayed at 0.10-0.12 per 1,000 characters through 2021-2024, then rose to 0.217 in 2025, more than double the baseline
- U.S. press writing follows AP style with spaced dashes, making the unspaced form a marker of LLM-assisted text
- The preregistered study is reproducible with public data
Why Is the Em-Dash an AI Fingerprint?
Large language models leave subtle stylistic traces in text written with their help, and the most discussed is the em-dash (Unicode U+2014), especially the unspaced form word—word. In typeset English prose, the unspaced em-dash is common, but U.S. press writing follows AP style, which calls for spaced dashes. So when an official U.S. press release is full of unspaced em-dashes, it is a fairly clear sign that AI helped write it.
The Data: A Sudden Doubling in 2025
Using a preregistered design (OSF: 10.17605/OSF.IO/U5NEY), the study drew 146,239 releases from 480 congressional offices via the open congress-press dataset, measuring the density of unspaced prose-form em-dashes per 1,000 characters of cleaned text, with Poisson/negative-binomial models, a length offset and clustering by office. Results show density held at 0.10-0.12 per 1,000 characters through 2021-2024, then rose to 0.217 in 2025, more than twice the baseline; the share of releases containing such em-dashes also increased.
What It Means for Tech Writing
The significance goes beyond political communication: it offers a quantifiable method for detecting how deeply LLM-assisted writing has penetrated large text corpora. For tech media it hits close to home. Many newsrooms, including ours, explicitly ban the em-dash in style guides precisely because it is the most visible AI writing habit. When a stylistic preference becomes a measurable scientific indicator, we gain an objective yardstick for how widespread AI writing has become.
Source: arXiv — The em-dash em-beds in Congress (2608.05889)
