Evidence and policy
Prosodeus measures repeated phrases and repetitive sentence structure. It cannot tell you who or what wrote a text, so don't use it to decide that.
Where the word lists come from
Kobak et al. (Science Advances, 2025) compared word frequencies across roughly 15 million PubMed abstracts from 2010 to 2024. They found 379 words whose use jumped in 2024. Most were style words like "delves". Few had anything to do with medicine. Most of the vocabulary categories in the hit lists are based on this study.
Detector accuracy
- OpenAI withdrew its own AI-text classifier in July 2023 for low accuracy. In its launch test it caught 26% of AI-written text and flagged 9% of human-written text by mistake.
- Liang et al. (Patterns, 2023) tested seven widely used GPT detectors on TOEFL essays by non-native English speakers. On average the detectors misclassified 61.22% of them as AI-generated.
- Weber-Wulff et al. (2023) tested a range of detection tools and found them inaccurate and unreliable.
Models learned every phrase on these lists from human writing. The lists also flag non-native writers, formal academic prose and anyone who runs a grammar checker. Treat a hit as a remark about style. A student accused on the strength of a flag can lose a grade over a phrase that plenty of people write.
Where the evidence is weaker
- We built the 1,000-phrase dataset with LLM-assisted research, and nobody has re-checked every entry by hand. Each entry records its signal strength, its false-positive risk and where the claim came from.
- Don't quote statistics from those research runs. The figures on this page come from the published papers.
- The lists will go out of date as models change and as writers drop words they know get flagged. Sentences that make no checkable claim will be around for longer, and the rewrite guidance is aimed at those.
- The model-specific category lists habits associated with particular model families. The families are trained in similar ways and share most of their habits, so the category cannot tell you which model wrote a text.
How the tiers work
| Tier | Rule | How to treat it |
|---|---|---|
strong | Strong contextual signal and low or medium false-positive risk | Can be flagged alone. More telling next to other hits |
moderate | Moderate signal, or strong signal with other caveats | Flag only alongside other hits |
watch | Weak signal or high false-positive risk | Common in ordinary writing. Never flag it alone |
Tiers are labels for human readers and for the prompts given to models. The analysis engine ignores them when it scores a document. As a rough guide, check any paragraph of about 100 words with two or more markers. At five or more, the paragraph usually needs rebuilding.