The em-dash em-beds in Congress: A population-level rise in em-dash frequency in U.S. congressional press releases at the dawn of the large-language-model era, 2021-2025

Author: Przemysław Czuma, MD, MBA (Polish Association for Artificial Intelligence in Medicine; ORCID 0009-0009-8235-2053)

The study examines whether a stylistic trace of large language model (LLM) use — the em-dash (—), especially the unspaced form word—word — can be measured in U.S. congressional press releases. The author notes: "LLMs have no keyboard cost. They can produce an em-dash as easily as any other token and often use the unspaced English form word—word. In recent years, this mark has become an informal sign that readers associate with LLM-generated prose. The unspaced form is normal in typeset English prose but unusual in U.S. press writing, where AP style calls for spaces around a dash. The research question is deliberately narrow: did the frequency of one punctuation mark change, and when?"

The study was preregistered in the Open Science Framework (OSF; osf.io/u5ney; DOI 10.17605/OSF.IO/U5NEY) before any confirmatory result was calculated. The corpus was the open congress-press dataset of full-text press releases from members of the U.S. House and Senate. To avoid a source-composition seam, the analysis used only scraper-sourced records from 2021–2025 (N = 146,239; 480 offices). The primary endpoint was the density of unspaced prose-form em-dashes per 1,000 characters of cleaned text, modeled with Poisson/negative-binomial regression, with text length as the exposure and uncertainty adjusted for repeated releases from the same office (Bioguide). The preregistered cut-off was the public release of ChatGPT on 30 November 2022, approximated by calendar years (pre = 2021–2022, post = 2023–2025). Mandatory boilerplate removal targeted header datelines CITY, ST —, which would otherwise distort the measure. The prespecified effect-size thresholds required a ratio of at least 1.5× and an absolute increase of at least 0.05 per 1,000 characters. A validation (STOP/GO) gate checked dateline-removal fidelity, and preregistered placebo cut-offs (June 2021, January 2022, June 2022) and an ASCII-hyphen net-new control were specified.

Main result. Prose-form em-dash density was 0.102 per 1,000 characters in 2021–2022 and 0.158 in 2023–2025. The exposure-weighted primary frequency ratio was 1.55 (95% CI 1.28–1.93), just above the 1.5× threshold, and the absolute increase was +0.056 per 1,000 characters; with the exact daily cut-off it was +0.054. Both exceed the prespecified +0.05 threshold.

Timing. The annual series showed stability followed by a sharp 2025 rise: density was 0.0997 in 2021, 0.1035 in 2022, 0.1047 in 2023, and 0.1157 in 2024. It then rose to 0.2167 in 2025. That is about twice the four-year baseline (×2.04). Prevalence rose from about 13% in 2021–2023 to 24.8% in 2025. Quarterly data sharpened the timing: Across the sixteen quarters of 2021–2024, density stayed around 0.08–0.13 per 1,000... It then rose from 0.130 in 2024Q4 to 0.193 in 2025Q1 and 0.241 in 2025Q2. About 97% of the total increase occurs in 2025. An interrupted time-series model showed "a flat pre-cut-off trend (×1.01/year)... no upward step at the cut-off (step ratio 0.76)... a post-cut-off slope of ×1.45/year... In other words, the model supports acceleration after the cut-off, not an immediate upward step at it."

Controls and falsification. The increase was net-new: ASCII-hyphen density stayed stable: 2.75 / 2.58 / 2.79 / 2.76 / 2.68 per 1,000 characters from 2021 through 2025, so conversion from hyphens explains 0% of the increase. It appeared within authors: Among 262 offices with at least 10 releases in both calendar eras, 75.6% increased their em-dash density; the median change was +0.036 per 1,000, and the office-level sign test gave p ≈ 1×10−16. A closed panel of 224 offices present every year showed 0.100 / 0.091 / 0.107 / 0.103 / 0.192 from 2021 through 2025, meaning the same offices changed style in 2025. The result was not driven by prolific offices (removing the top 5% left the ratio at 1.66). The three placebo cut-offs inside the pre-LLM window were all flat (RRs 0.989, 1.058, 0.941). The spaced form also rose roughly twofold in 2025, so the 2025 discontinuity appears in both spaced and unspaced forms.

The full preregistered decision rule was not met as written. The manual adjudication of 300 releases found stripping recall was 99.2% (122/123), but the automatic residual-leak rate was 7.00 / 4.89 / 3.94 / 4.95 / 4.94% per year in 2021–2025 (weighted mean 4.96%), above the preregistered 2% threshold, and its change exceeded the registered 1-point threshold. The author emphasizes: the validation gate was formally breached, so the study's full preregistered decision rule was not met as written. The interpretation is therefore exploratory rather than confirmatory. Additional deviations included the era aggregation choice (pooled estimate 1.55× passes; simple yearly average gives 1.43× and fails both thresholds), the calendar-year boundary approximation (exact daily cut-off gives 1.528 and +0.054), and several post-registration analyses (pipeline invariance check, continuing-vs-new offices, party/chamber stratification, spaced-form analysis, lexical analysis).

The author concludes the rise was broad, appeared within offices, and was not explained by the measurement artifacts tested here. The result is a population-level indicator, not an authorship detector for individual releases — a human writer may use many em-dashes, and an LLM user may remove every one. The timing supports a delayed-diffusion interpretation: public availability in 2022 was a necessary early condition, while later improvements made assisted drafting, editing, shortening, and translation practical enough for regular use. A key competing explanation is noted: January 2025 was also the start of the 119th Congress... Staff turnover or a new political communication climate could therefore change writing style without LLM diffusion. However, exploratory stratification showed the increase was similar across parties and chambers, with a within-office Republican/Democrat fold-change ratio of 0.96 (95% CI 0.67–1.38), the opposite of a simple story in which the incoming majority alone drives the change. The author ultimately frames the finding modestly: the style of official political communication changed measurably and quickly, consistent with LLM-assisted writing diffusion, but causality and individual authorship cannot be established from this corpus-level marker.

1. Population-level AI-text auditor

Build a system that monitors punctuation frequency (especially unspaced em-dash density per 1,000 characters) across large text corpora over time. It would use exposure-weighted Poisson/negative-binomial models, cluster errors by source (e.g., congressional office), and apply preregistered cut-offs and placebo dates. The system can detect a ≥1.5× frequency ratio and ≥0.05 per-1,000 absolute increase as an early indicator of LLM-assisted writing adoption, while avoiding per-document false accusations.

2. Within-author style-change detector

Create an AI system that compares the same author/organization's output across eras using closed panels and sign tests. It can identify whether a style shift is truly broad-based (e.g., 75.6% of offices increasing em-dash density) rather than driven by a few prolific generators. This improves detection of genuine LLM diffusion versus compositional changes or new-office effects.

3. Falsification-aware text analytics

Improve AI audit systems by automatically testing competing explanations: stable ASCII-hyphen density (to rule out punctuation conversion), spaced-dash control (to rule out a single-form artifact), and placebo cut-offs inside the pre-LLM window (to rule out unrelated secular trends). This yields more robust conclusions about whether a stylistic discontinuity is likely associated with LLM availability.

4. Genre-aware LLM writing assistants

Enhance LLM text generators with stylistic calibration for publishing norms. For U.S. government/press-release writing, the assistant can learn AP-style spaced dashes and match historical punctuation distributions, preventing the unintentional LLM fingerprint of unspaced em-dashes. The improved system can adjust decoding or post-edit punctuation to keep frequency within a corpus-specific baseline.

5. Interrupted time-series adoption estimator

Build an AI system that fits an interrupted time-series model to punctuation/style metrics around known LLM release dates (e.g., ChatGPT, 30 Nov 2022). It can distinguish an immediate step at release from a delayed acceleration (slope change after release), providing a more accurate picture of how quickly AI writing tools diffuse into real-world official communications.

6. Preregistration and validation-gate tooling

Create a statistical workflow that requires a validation gate (e.g., residual dateline leakage ≤2%) before confirmatory results are accepted. If the gate is breached, the system automatically downgrades findings from confirmatory to exploratory, reducing false confidence. This improves reproducibility and honesty of AI-driven text-forensics studies.

7. Robust multi-threshold decision system

Improve AI detectors by using combined thresholds—rate ratio ≥1.5× and absolute increase ≥0.05 per 1,000 characters—rather than relying on a single metric. The system can also report sensitivity to year aggregation (e.g., pooled ratio vs. simple yearly average) so users can see how fragile a claimed effect is.

8. Lexical-punctuation joint model

Build an AI system that jointly models punctuation and vocabulary shifts, not just one marker. The paper's exploratory lexical analysis suggests that combining em-dash density with word-level changes would improve detection of LLM-assisted editing, while remaining transparent that the result is a population-level indicator, not an individual authorship detector.