Evidence
Style can be a signal. Style alone is not proof of AI authorship.
This page collects concise evidence used by Em Dash Rights. It is not a claim that AI-authorship detection never works, and it is not a claim that stylistic patterns are unrelated to generative AI.
The distinction is narrower: a population-level statistical observation is not the same thing as an individual determination of authorship.
97%
In a 2023 Stanford-led study, 89 of 91 human-written TOEFL essays were flagged as AI-generated by at least one of seven detectors tested.
Stanford HAI / Patterns · 2023
This describes that study’s result on that set of essays. It is not a claim that AI detectors have a 97% false-positive rate in general.
<20%
Turnitin treats AI-writing results below 20% differently because false positives are more likely in this range and does not report an exact percentage.
Turnitin AI Writing Report
This does not mean every result below 20% is incorrect. It means Turnitin treats that range as less reliable for an exact percentage.
69,632
A 2026 preregistered study examined 69,632 scientific preprints and found an increase in em-dash use in the generative-AI era.
Preregistered research · 2026
The finding is useful at a population level. It does not make an em dash a per-document authorship detector.
Source labels name the origin of each claim. A source is not linked until a stable public URL is recorded in the evidence catalogue. Missing links are not replaced with guessed citations.
This page will expand with methodology notes and further citations. The standard will not:
