AI-generated text leaves measurable statistical fingerprints that differ from human writing. Our tool measures 16 statistical properties of your text. Only some of them count toward your score - we tested each one against papers written before ChatGPT existed, and dropped the ones that were wrong:
- Vocabulary diversity (3 metrics): Lexical Diversity (MTLD) measures how many words you can read before vocabulary starts repeating; Hapax Legomena tracks how many words appear only once; Yule’s K is a classical statistical measure of vocabulary repetitiveness. These are not used in your score. We expected AI to repeat words more than people; measured against verified human papers, the opposite was true - real academic writing repeats its key terms constantly, while AI varies its wording. Scoring these would flag careful and non-native writers for no reason, so they carry no weight.
- Sentence rhythm (4 metrics): Sentence Length Uniformity (coefficient of variation), Sentence Length Variation (raw standard deviation), Burstiness (alternation between simple and complex passages), and Sentence Opening Diversity (whether you keep starting sentences with the same words). AI keeps sentence lengths and openings unnaturally consistent.
- Phrase fingerprints (2 metrics):AI-Characteristic Phrases checks for 60+ expressions LLMs over-favor - “it is important to note,” “delve,” “multifaceted,” “tapestry,” “leverage,” etc. Transition Word Frequency catches sentence-starters like “Furthermore,” “Moreover,” and “Additionally,” which AI uses two to three times more often than typical academic prose.
- Document structure (3 metrics): Paragraph Length Uniformity flags suspiciously even paragraph sizes; Passive Voice Usage spots academic-context overuse; Adjacent Sentence Similarity measures whether sentence-to-sentence flow is too smooth - AI tends to produce uniform transitions, while human writing has more varied connections.
- Statistical fingerprints (4 metrics): Zipf Exponent compares your word-frequency curve to the natural-language norm; Character-Level Entropy detects predictability at the letter level; Punctuation Diversity counts how many distinct punctuation types you use; Word Length Variation captures the spread of short vs. long words. These are subtle signals individually but combine into a strong fingerprint.
Each metric is scored 0 - 100 and weighted (heaviest on AI phrases, sentence rhythm, and lexical diversity) to produce one overall AI probability score. We follow the methodology from Desaire et al. 2023 and the Tercon 2025 survey. This is a heuristic analysis - no AI detector is 100% accurate, but the per-metric breakdown shows you exactly which patterns triggered each score so you can revise specifically rather than guessing.
When you run the cleaner on a result, we apply the same rule set in reverse: AI-favored phrases get swapped for plainer alternatives and overused sentence-starting transitions are removed. For DOCX and PDF uploads, the formatting of your original document is preserved - only the flagged words are changed in place - and the “Changes made” panel shows every replacement as deleted text in red and added text in green so you can review them before submitting your work.