LLM-Style Em-Dash Spike in Congress Press Releases (2025)

LLM-Style Em-Dash Spike in Congress Press Releases (2025)

When you skim a political press release, punctuation usually feels like background noise. But new research published as a preprint by Przemysław Czuma (see arXiv:2608.05889) suggests one tiny punctuation choice—the em-dash in the unspaced “word—word” form—might have become more common in U.S. congressional messaging right as the large-language-model era matured.

The punchline is striking: across 146,239 congressional press releases from 2021–2025 (from 480 offices), the frequency of unspaced prose-form em-dashes stayed remarkably steady through 2024, then more than doubled in 2025. And the analysis was designed to catch common “data artifact” explanations—like changes in extraction formatting—so this doesn’t look like a simple scraping glitch.

Still, the story is nuanced. The study’s main “confirmatory” decision rule was slightly breached due to a validation gate issue, so the author treats the result as exploratory-but-strong rather than fully “case closed.” Even so, the pattern is broad, shows up within the same offices, and arrives in 2025 in a way that fits how LLM diffusion might unfold in real workplaces.

Why This Matters: The Real-World Signal Isn’t “Who Wrote It,” It’s “How They’re Writing Now”

This research is significant right now because AI-assisted writing is no longer just a research curiosity—it’s operational. Teams are adopting tools, workflows, and editing habits, and those habits can leave statistical fingerprints in documents that are otherwise hard to audit. The em-dash is a weird little candidate for “AI trace detection,” but that’s exactly why it’s interesting: it’s not a high-level keyword or ideology marker. It’s a micro-style choice that shows up in the mechanical layer of writing.

A practical scenario where this could be useful today: imagine a newsroom, a civil society group, or even a congressional oversight shop trying to monitor whether official communications are becoming stylistically homogenized—or whether a particular office’s output suddenly shifts without an obvious human explanation. You wouldn’t use this to call any single release “AI-written.” Instead, you’d track population-level style shifts over time, the way you might track changes in data formatting or translation patterns.

Where this builds on earlier AI research is also important. Prior work has looked at signs of AI modification in many domains (papers, reviews, other corpora). This study does something different: it asks whether a specific, document-genre-friendly marker is measurable in a setting where typography tends to be preserved—official congressional press releases—and then tests whether the trend is plausibly caused by text extraction artifacts versus real writing changes. That’s the kind of “make it measurable, then sanity-check it” approach the field needs as AI signals become more subtle.

What the Researchers Actually Measured: Unspaced “word—word” Em-Dashes Per 1,000 Characters

The study’s central idea is straightforward: an em-dash can be written in different ways, and one form is unusual in U.S. press writing.

• In typeset prose, people often write the em-dash in the unspaced form: word—word
• But U.S. press writing commonly follows AP style, which calls for spaces around a dash (so you’d expect something like word — word)

Czuma focuses on the Unicode em-dash character U+2014, and then classifies each em-dash instance into categories. The primary signal is the prose-form unspaced pattern (think word—word). The study excludes other types (like numeric ranges or capitalized term relations) because those are less likely to be the “LLM-style” trace the author is looking for.

The core endpoint: density, not just presence

Instead of asking “does this release contain an em-dash,” the author counts density:

• Unspaced prose-form em-dashes per 1,000 characters
• This matters because long releases naturally have more punctuation opportunities.
• Density also gives a cleaner comparison across offices and years.

The dataset size and what counts as an “office”

The corpus is from the open congress-press dataset (House + Senate press releases). The analysis uses only scraper-sourced records for 2021–2025 to avoid a “source seam” (the dataset has an older API import and a newer scraper pipeline with different typography).

Key numbers:
- N = 146,239 releases
- 480 offices across 2021–2025
- An “office” is identified using the Bioguide member ID (so repeated releases from the same office are treated as clustered for uncertainty estimates)

And one detail that shows the author is thinking like an auditor: the study doesn’t just count punctuation. It includes a boilerplate-removal step (more on that below) so the measurement targets the body text rather than the template.

The pre-LLM vs post-LLM split is anchored to ChatGPT’s public release

The preregistration (OSF: osf.io/u5ney, DOI: 10.17605/OSF.IO/U5NEY) used an exposure cutoff tied to:

• ChatGPT’s public release date: 30 November 2022

For the main manuscript analysis, the author approximates this with calendar-year eras:
- Pre: 2021–2022
- Post: 2023–2025

(There’s also a sensitivity analysis using the exact daily cutoff, which keeps the conclusion intact, but the main results still treat this calendar approximation as a deliberate modeling choice.)

How the Study Prevents “It’s Just Formatting” From Becoming a False AI Story

Whenever someone tries to read AI traces from text, you have to fight two enemies:

1. Typography/extraction artifacts (scrapers change output; templates change)
2. Sampling/composition artifacts (which offices you see in the dataset changes over time)

This paper is unusually careful about both.

A “boilerplate gate” because congressional headers always contain a dash

Every congressional press release includes a header line that looks like:

• CITY, ST —

That header always includes an em-dash, but it’s part of the dateline template, not the writer’s prose. If you didn’t remove it, you’d measure template punctuation rather than office style.

So the author applies a fixed removal rule:
- Remove the dateline only when the state token matches a closed list of 50 states, DC, and territories
- This conservative approach avoids accidentally cutting a sentence that just happens to start with something like Word — ...

A net-new control: ASCII hyphen stability

If em-dashes “spike” because of a conversion bug—say hyphens turning into em-dashes—you’d expect the opposite to happen to hyphens.

That’s why the study uses ASCII hyphen density (U+002D) as a control:
- The em-dash rise is interpreted as net-new only if hyphen density stays flat rather than dropping.

A comparison across offices: within-author change

Another classic failure mode is that a result is driven by a few offices, or by offices that enter late.

To address this, the study:
- models repeated releases with clustering by office (member ID)
- and also checks within-office change directly:
- among offices with at least 10 releases in both eras, 75.6% increased their em-dash density

Artifact checks that specifically target the 2024/2025 boundary

A single-year jump could come from an extraction pipeline update. The author tests pipeline invariance:
- If the scraper behavior changed at the boundary, other date/header-related metrics should show a step.
- The checks did not show a 2024/2025 seam in the dominant scraping signals.

But there’s an important caveat: the validation gate was formally breached

In a preregistered workflow, there’s supposed to be a “STOP/GO gate” to decide whether the measurement is trustworthy enough to make the confirmatory claim.

Here, the author reports that the full preregistered decision rule was not met as written due to a residual-dateline-leak threshold. That doesn’t necessarily invalidate the trend (the paper includes sensitivity analyses arguing it doesn’t explain the increase), but it does change how confidently the result should be interpreted: more “exploratory with checks” than “fully confirmatory.”

The Big Result: Stable Until 2024, Then a 2025 Em-Dash Doubling

Let’s talk numbers.

Main finding (pooled/model-era comparison)

Across the whole corpus:

• 2021–2022 density: 0.102 unspaced prose-form em-dashes per 1,000 characters
• 2023–2025 density: 0.158 per 1,000 characters
• Frequency ratio (post / pre): 1.55 (95% CI 1.28–1.93)

That ratio clears the preregistered practical threshold of 1.5×—barely, but clearly. The absolute increase is also reported:

• around +0.054 to +0.056 per 1,000 characters (depending on how the cutoff is treated)

The study also reports prevalence:
- the share of releases containing at least one unspaced prose-form em-dash rises to 24.8% in 2025
- from about 13% during the stable baseline years (roughly 2021–2023)

The timing is “sudden acceleration” in 2025

Annual densities:
- 2021: 0.0997
- 2022: 0.1035
- 2023: 0.1047
- 2024: 0.1157
- 2025: 0.2167

That’s about a 2.04× jump relative to the four-year baseline.

Quarterly data sharpens it further:
- almost 97% of the total increase happens in 2025
- density rises from 0.130 (2024Q4) to 0.193 (2025Q1) and 0.241 (2025Q2)

The control stays flat: hyphens don’t disappear

ASCII hyphen density stays near 2.7 per 1,000 characters across 2021–2025, which supports the interpretation that this is more em-dash usage, not just a conversion artifact.

The increase isn’t just “new offices”

In a continuing vs new offices check:
- offices already present every year still jump from 0.103 in 2024 to 0.192 in 2025
- offices newly covered in 2025 are higher, but they don’t explain the whole rise

Within offices, most offices change

In the within-author analysis:
- among 262 offices with ≥10 releases in both calendar eras
- 75.6% increased their em-dash density
- median office-level change: +0.036 per 1,000
- sign test p is reported as about 1×10⁻¹⁶

That broad “most offices moved” pattern is what makes the signal hard to dismiss as a one-off glitch.

A compact comparison of the era cut

Here’s how the paper frames the main effect and the key controls:

Why the Timing Could Make Sense—and What Competing Explanations Still Exist

The million-dollar question is: did LLM adoption cause this? The study can’t directly prove causation. It measures a style trace at the population level.

But it does consider interpretations that fit the timing.

Delayed diffusion: “not when LLMs arrived, but when they became practical”

Czuma discusses a delayed diffusion hypothesis: models became publicly available in late 2022, but teams might not adopt them for routine drafting until later when they were more reliable, less verbose, and more natural for institutional work.

That would produce:
- stable style early on
- change emerging later
- a surge in 2025 when tool use becomes “good enough” for regular workflows

Competing explanation: 2025 also starts a new Congress

There’s also a very real political-world confound:
- January 2025 marks the start of the 119th Congress
- with shifts in administration and Senate majority (depending on the timeline framing in the paper)

Staff turnover and communication style changes after elections could plausibly alter punctuation habits even without LLM tools.

The paper partially checks this:
- within-office changes appear for both parties
- the stratified patterns are described as broadly symmetric in a way that argues against a simple “new majority did it” story

However, because this is still a single corpus and a single genre, the study can’t fully separate “LLM diffusion” from “post-election staff and style churn.”

A key limitation: the em-dash isn’t an individual-release detector

Even if the population-level trace is real, the em-dash density cannot tell you:
- which specific release was written by an LLM
- whether a particular office adopted LLMs or just changed house style

So the findings are best treated as an instrument for monitoring rather than an authorship microscope.

Key Takeaways

• A measurable punctuation trace exists: In U.S. congressional press releases, the unspaced prose-form em-dash (word—word) rose sharply in 2025.
• The pattern is big and broad: Density jumped from about 0.10–0.12 (2021–2024) to 0.2167 in 2025, and the share of releases containing at least one such em-dash rose to 24.8%.
• Controls suggest it’s net-new writing style, not formatting conversion: ASCII hyphen density stayed stable across 2021–2025.
• Within-office evidence supports diffusion rather than random sampling: 75.6% of offices with enough releases increased their em-dash density.
• Timing matters: The increase is concentrated in 2025 (about 97% of the total increase happens then), which fits plausible “delayed diffusion” narratives for LLM adoption.
• Interpretation is exploratory, not fully confirmatory: The preregistered validation gate was formally breached due to a residual dateline leak threshold, though sensitivity analyses argue it’s unlikely to explain the full trend.
• Practical use case today: You could use similar “style fingerprint” monitoring to detect large-scale shifts in official communication—useful for transparency and research—without claiming to identify AI use in any single document.
• What’s still unknown: This doesn’t prove LLMs caused the change, and it can’t separate tool adoption from other 2025-related changes like staff turnover.

If you want to dig deeper, the full study is available at arXiv:2608.05889, with preregistration details linked via OSF.