Congress's press releases doubled their em dashes. Two years after ChatGPT.

Right, the court is in session. The defendant is a punctuation mark, the em dash, and the charge is that it spread through congressional press releases once language models arrived. I’m only the court reporter, so I take no side. I just write down what was said.

The evidence comes from a single-author preprint by Przemysław Czuma, of the Polish Association for Artificial Intelligence in Medicine, titled “The em-dash em-beds in Congress”. It’s from August 2026, it isn’t peer reviewed, but it was preregistered on the Open Science Framework before any confirmatory result was calculated. Fair point in its favour.

Why press releases? Well, AP style puts spaces around a dash, so the unspaced form written straight between two words is unusual in this genre. Low human baseline. Member offices also publish without an outside typesetter, so their own typography survives.

The study took 146,239 releases from 480 House and Senate offices, 2021 to 2025, and counted unspaced em dashes per 1,000 characters of cleaned text, after stripping the “CITY, ST” dateline. The preregistered cut-off was ChatGPT’s public release on 30 November 2022. The uncertainty was adjusted for repeated releases from the same office, and three placebo cut-offs were tested.

The prosecution goes first. Density sat between 0.10 and 0.12 from 2021 through 2024, then rose to 0.217 in 2025, more than twice the four-year baseline. The share of releases with at least one such em dash went from about 13% to 24.8%. The primary ratio was 1.55 (95% CI 1.28 to 1.93), just above the preregistered 1.5 times threshold. Of 262 offices with enough releases in both periods, 75.6% increased their use, and a closed panel of 224 offices went from 0.103 in 2024 to 0.192 in 2025. Same offices, changed style. Dropping the 24 busiest offices left the ratio at 1.66, so a few prolific offices aren’t carrying it. Hyphens stayed near 2.7 per 1,000 characters, the en dash stayed flat, and the placebo cut-offs showed no increase.

Now the defence, and it’s a good one. Nothing happened when ChatGPT arrived. Density was 0.1047 in 2023 and 0.1157 in 2024. The time-series model found no upward step at the cut-off, “if anything a temporary dip”. About 97% of the increase falls in 2025. The author writes that the change “appeared substantially later than the simplest ‘ChatGPT released, writing changed immediately’ story would predict.”

Then there’s January 2025: a new Congress, a new administration, a change in the Senate majority. Staff turnover could shift style without any model. The rise was about 1.9 times for both Democrats and Republicans within continuing offices, which weighs against a simple party story, but the author says the corpus “cannot fully separate” the two.

The quality gate also failed. The leftover dateline rate was above the preregistered 2% threshold, so “the full preregistered decision rule was not met”. Hence the LLM reading is “exploratory rather than confirmatory”. Averaging yearly figures instead of pooling them gives 1.43 times, which fails the thresholds.

I should say plainly what this measures: one mark becoming more common, its prevalence. It doesn’t say who wrote any release, and it doesn’t show that LLMs caused anything. The author agrees: “the em-dash is not a detector”, and the study “does not support a causal claim”. One country, one genre.

So the mark really did more than double, broadly and within the same offices. But it didn’t move when ChatGPT arrived, the jump coincides with a new Congress, and the author calls the explanation exploratory.

Verdict: split.