Guide — a heuristic check for AI-written text
12 weighted writing signals, a sentence heatmap, a score with a likely range, rewrite tips and compare. Runs in your browser, no AI model, no upload.Paste into the INPUT pane, press paste, open file or drop a TXT, DOCX, ODT, PDF or HTML file. Files are read in your browser. Use at least 100 words.
With Auto-analyze on, the score updates as you type; otherwise press ▶ analyze or Ctrl/⌘ + Enter. Sensitivity switches between Conservative, Medium and Aggressive.
The ring shows the AI likelihood and, in a faint band, the likely range. Short or mixed texts get a wide range. The Looks AI / Looks human lists explain the score.
The heatmap colours every sentence from green to red. Hover for the reasons; click a bar in the Sentence Spectrum or a sentence in Most AI-like to jump to it.
tips lists concrete rewrites. 📌 pin up to 4 texts and open compare. Copy or download a report, or copy a 🔗 share link.
Quick reference
- 0–24% / 25–44%
- Likely Human / Probably Human
- 45–64%
- Uncertain / Mixed
- 65–84% / 85%+
- Likely AI-Generated / Very Likely AI
- 20 / 13 / 13%
- AI vocabulary · sentence variety · personal voice
- 10 / 8 / 7%
- templates · transitions · punctuation
- 6 / 6 / 5%
- paragraphs · openers · repetition
- 5 / 4 / 3%
- sentence length · passive · vocab richness
- green → red
- human-like → AI-like
- phrase marks
- red strong · purple template · orange transition · yellow word
- hover
- + reasons pushing up, − reasons pushing down
- length
- under ~150 words the score is pulled toward the middle
- language
- phrase lists are English
- method
- hand-tuned heuristics, no trained model, no measured accuracy
AI text detection, explained
This detector does not ask an AI whether it wrote the text. It measures writing patterns that are common in language-model output, such as model vocabulary, even sentence lengths, rhetorical templates and a missing personal voice, and combines them into one score with a likely range. That makes it fast and private, but also easy to fool and prone to false positives.
How does this AI text detector work?
The text is split into sentences and paragraphs and scored on 12 heuristic signals:
| Signal | Weight | What it measures |
|---|---|---|
| AI vocabulary | 20% | Words and phrases that models overuse ("delve", "tapestry", "plays a crucial role"), per 100 words; strong phrases count 2.5× |
| Sentence variety | 13% | Burstiness: the coefficient of variation of sentence lengths |
| Personal voice | 13% | Missing first person, contractions, concrete numbers and informal words |
| Rhetorical templates | 10% | "It isn't just X — it's Y", "not only … but also", "Here's why", "The result?", "Whether you're…", rule-of-three lists, wrap-ups and assistant sign-offs |
| Transitions | 8% | Furthermore, Moreover, Additionally and similar, per sentence |
| Punctuation profile | 7% | Many em dashes and colons, few exclamation marks, questions, brackets or ellipses |
| Paragraph uniformity | 6% | Paragraphs of near-identical length |
| Sentence openers | 6% | AI-typical starts and the same first word over and over |
| Self-repetition | 5% | Repeated three-word phrases |
| Sentence length | 5% | An average in the 15–26 word band |
| Passive voice | 4% | Share of passive constructions |
| Vocabulary richness | 3% | Moving-average type-token ratio (MATTR), which does not drop with text length |
Signals that need more text (for example paragraph uniformity needs three paragraphs) are left out and the other weights are rescaled. The weighted sum goes through a logistic curve to give the 0–100% score. No machine learning model or external API is used — all analysis runs locally in your browser.
What is the likely range?
One number suggests more precision than a heuristic has. The range around the score widens for short texts and when the signals disagree, and narrows for long texts where most signals point the same way. If the range crosses the 45–65% band, treat the result as undecided.
What does the sentence heatmap show?
Every sentence gets its own score from the markers it contains: strong phrases, model vocabulary, transitions, templates, an AI-typical opener, a rule-of-three list, an em dash or a length close to the average push it up; first person, contractions, numbers, informal words, questions, brackets and short fragments push it down. Scores are smoothed with the neighbouring sentences, because AI help usually covers a passage rather than a single sentence. That makes it easy to spot a pasted AI paragraph inside human writing.
What AI writing styles does this detect?
The phrase lists and templates are hand-picked from patterns often seen in output from models such as ChatGPT, Claude, Gemini, Llama and Mistral, including the newer chat style with em dashes, "not just X, it's Y" contrasts and rhetorical questions. They were not trained or benchmarked on a labelled test set.
The detector works best on English text. It may give unreliable results for other languages, highly technical text, poetry, dialogue, lists or very short passages.
How accurate is the AI detection?
Heuristic AI detection without a trained model is inherently probabilistic. This tool will produce false positives (human text scored as AI) and false negatives (AI text scored as human). Formal, academic and corporate writing and text by non-native speakers can score high; AI text that someone edited or asked to "sound human" can score low. There is no published accuracy figure for this tool. Use it as one signal among many, and never use the score on its own to accuse someone of using AI.
Which files can I open?
TXT, Markdown, CSV, HTML, DOCX, ODT and text-based PDF. Files are read in your browser; for DOCX/ODT and PDF the page loads the open-source JSZip or PDF.js library from cdnjs the first time you open such a file. Scanned PDFs contain images, not text, and cannot be read.
Is my text stored or sent anywhere?
No. All analysis runs in your browser. History is off by default; if you turn it on, the last 12 texts are kept in this browser's local storage only, and turning it off deletes them. A share link puts the compressed text after the # in the URL; browsers never send that part to the server, but anyone with the link can read the text. The page loads the privacy-friendly Simple Analytics script, which counts visits and never receives your text.
What is "sentence burstiness" and why does it matter?
Burstiness is the variance in sentence lengths. Human writers alternate short, punchy sentences with long ones. Models tend to produce evenly sized sentences. The detector uses the coefficient of variation (standard deviation ÷ mean); human prose is usually above 0.5, model output often below 0.35. The Sentence Spectrum chart shows it at a glance: an even row of bars is a warning sign.
How do the writing tips work?
The tips tab turns the strongest signals into concrete edits: plain replacements for flagged words ("leverage" → use, "delve into" → look at), the templates to drop, a run of same-length sentences to break up, transitions to cut and places to add your own example or number. They are writing advice for clearer, more personal text — not a guarantee of any score in another detector.
What is the minimum text length for reliable results?
At least 100 words, better 200 or more. Below about 150 words the score is pulled toward the middle and the likely range gets wider, and several signals switch off until there is enough text.