Rival Research · Stylometrics
The Em-Dash
Civil War
Every model, by em-dash use
Each dot is one model's mean em-dash rate across 43 standardized tasks. 26 models use no em-dashes at all; a long heavy tail sits opposite. There is no "average" model here. There are two populations.
The headline
The gap is widening, fast
For every task, we measure how much models disagree on a feature, the cross-model spread. If writing were converging, it would shrink. For the em-dash it more than quadrupled in a single year.
growth in the cross-model spread of em-dash use on identical tasks, from 2025-Q1 to 2026-Q1. The single most divergent feature in the entire fingerprint.
Within-task stddev: 0.61 → 2.50 em-dashes / 100 words.
Em-dash disagreement, over time
As the band widens, the camps move apart. It opens early in 2025 and never closes.
Split at 1.0 em-dashes per 100 words, the trough in the distribution, the catalog cleaves into a heavy camp averaging 2.8 and a sparse camp averaging 0.38. The line cuts across labs and through them: OpenAI sits firmly heavy, Claude 2/3 and the DeepSeek line use zero.
The em-dash heavy camp
81
The em-dash sparse / free camp
101
Plot em-dash use against emoji use and the structure holds: the heavy camp spreads along its own axis, the sparse camp packs into the corner. Typographic personality is becoming a coordinate, not a consensus.
The em-dash is the loudest case, not the only one. Every feature below tilts up: models agree less in 2026-Q1 than they did in 2025-Q1.
Diverging features · within-task spread, 2025-Q1 → 2026-Q1
Paragraph length, inline code, emoji, semicolons and italics all fan out. The percentage is the real change in cross-model spread.
Convergence is not entirely a myth, it just lives somewhere narrower than the headlines claim. Models are tightening up on rhythm: sentence length, its variance, ellipses, transitions and exclamation marks. They breathe the same way. They just dress differently.
Converging features · within-task spread, 2025-Q1 → 2026-Q1
Models converge on pacing and rhythm while diverging on punctuation and formatting personality. "AI all sounds the same" is true about cadence and false about style.
Show your work
The convergence mirage
So where does the "everything is converging" story come from? From not controlling for the task. Compare models by their overall fingerprint and convergence looks dramatic. Compare them on the same task and almost all of it evaporates.
of the "AI is converging" signal is a measurement artifact of which tasks each model happened to answer, not a real shift in how models write.
Naive 29.3% convergence → 5.6% once you hold the task constant.
Convergence signal · raw vs task-controlled
comparing overall model fingerprints
same prompt, different model
Hold the task constant and 29.3% convergence collapses to 5.6%. Even that residual is fragile.
The honest caveat
We will not oversell our own residual. That remaining 5.6% controlled convergence is weak and suggestive: the endpoint confidence intervals overlap and the trend is not monotonic. We label it WEAK_SUGGESTIVE, not proven. The robust finding is the opposite of the popular one: feature-level divergence, led by the em-dash, is large, consistent and easy to see.
Method
How we measured this
The trap in any "is AI converging?" question is that the set of tasks each model answered changed over time. Compare a 2026 model and a 2023 model on different prompts and you measure the prompts, not the era.
So every cross-model comparison here is computed on the identical task. For each response we find the nearest response to the same prompt by a different model, using a 27-dimension, globally z-normalized stylometric vector and Euclidean distance. Responses are bucketed by the author model's release date.
Cite this
Rival (2026). The Em-Dash Civil War: AI models are diverging, not homogenizing. 183 models, 3,095 responses, 43 tasks. rival.tips/research/em-dash-civil-war
The Model Similarity Index
Which models write identically, and what it costs you.
Live model pricing
Input / output prices per 1M tokens, across every lab.
This report was written with zero em-dashes. We picked a side.