An analysis of Polish parliamentary questions shows that em-dashes cannot identify machine-written text, but subtle word frequencies shifted toward artificial intelligence profiles after 2023. Many observers assumed that long punctuation dashes would easily expose machine generated writing in any language. Instead, translation pipelines and local software insert shorter en-dashes, while the real footprint appears in the unconscious background distribution of conjunctions, prepositions, and particles.
When language models generate Polish text, their algorithms place the shorter en-dash mark on the page. These automated systems also balance closed-class words like prepositions much like an invisible metronome that human writers never intentionally control. As political offices paste or edit draft questions with software assistance, those mechanical word frequencies filter directly into the final submitted records. The documents show no sudden jump right after late 2022, but rather a gradual drift that solidifies across 2024 and 2025.
A preregistered investigation evaluated 60,275 single-author written parliamentary questions submitted to the Polish Sejm between 2015 and 2025. The test found that the proportion of documents with em-dashes rose by only 1.73 percentage points, failing the preregistered threshold because the confidence interval spanned negative 4.44 to positive 8.07 points. In contrast, the rank correlation of subconscious function words shifted positively to 0.368 against model profiles, spreading across 61 of 84 measurable parliamentary offices.
Researchers can now track large-scale shifts in writing tools across broad document archives without relying on flawed single-punctuation markers. This statistical method evaluates population-wide stylistic changes over time while leaving individual author identities and specific document origins unassigned.