METHODOLOGY · UPDATED JULY 24, 2026
How the UnAI.se writing-pattern score works
UnAI.se is a browser-local style analysis tool. It looks for patterns that often make text feel formulaic or machine-like and turns those observations into revision suggestions. It does not identify the author of a passage and cannot prove that text was generated by AI.
Human writing can score high and AI-assisted writing can score low. Do not use this result alone for grading, discipline, hiring, publication, plagiarism findings, or any other high-impact decision.
What the tool analyzes
The detector applies deterministic rules in your browser. The same text and settings should produce the same score in the same version of the tool. The five displayed dimensions summarize related groups of signals:
1. Marker words
Counts stock phrases, filler language, heavy hedging, and repeated expressions associated with generic prose.
2. Transitions
Looks for dense or predictable linking words such as “furthermore,” “additionally,” and “in conclusion.”
3. Sentence variation
Compares sentence lengths and flags unusually uniform rhythm or repeated syntactic shapes.
4. Pronoun patterns
Checks impersonal or repetitive constructions that can make prose feel distant and templated.
5. Structural regularity
Reviews repeated paragraph patterns, list-like progression, passive constructions, and other signs of rigid organization.
How the score should be read
The percentage is a normalized summary of the rules that fired. It is not a probability that AI wrote the text. A higher score means the passage contains more of the patterns this tool watches for; a lower score means fewer of those patterns were found.
- 0–29: fewer watched patterns. This does not certify human authorship.
- 30–59: mixed signals. Review individual highlights before changing anything.
- 60–100: many watched patterns. The text may benefit from revision, but authorship is still unknown.
Short text is especially unstable because a single word or sentence can change the percentage sharply. For a more useful result, analyze at least a full paragraph and interpret the highlighted evidence rather than the headline number alone.
Known limitations and false positives
Formal, technical, academic, legal, instructional, and second-language writing often uses regular structure and predictable transitions. These genres can score high even when a person wrote every word. Lists, templates, corporate style guides, and heavily edited copy can behave similarly.
The opposite is also possible. AI-assisted text may score low after manual editing, strong prompting, translation, or deliberate variation. New model styles can also differ from the patterns encoded in this version.
Because the system is rule-based, it does not compare text with a secret database, inspect document history, or access model provenance. It cannot detect plagiarism, factual accuracy, citations, intent, or whether AI use complied with a particular policy.
Privacy and local processing
The detector and rewrite rules run in your browser. Draft text is stored only in browser storage on your device so history and unfinished work can be restored. You can clear it from the tool or your browser. The project does not need an account and the current detector does not send draft text to a server. See the privacy policy for the full explanation.
Responsible workflow
- Read the highlighted sentence and the reason shown.
- Decide whether the pattern actually hurts clarity, accuracy, or voice.
- Rewrite only when the suggestion improves the passage.
- Recheck facts, quotations, citations, names, and numbers.
- Follow the disclosure and authorship rules that apply to your school, client, employer, or publisher.
For practical examples, continue with the plain-language guide to removing formulaic AI-sounding prose, or open the detector.