Text Summarizer
Picks the 1 to 8 sentences that matter most. Rule-based.
An in-browser estimate of how 'AI-like' text looks, based on sentence-length variation and vocabulary diversity. You can see the formula, and the result is a hint, never proof.
Your browser is preparing the tool. It runs 100% locally.
This AI Content Detector Helper estimates how 'AI-like' a passage looks using two open signals: how much your sentence lengths vary (burstiness) and how often words repeat (vocabulary variety). It combines them into a 0-to-100 score and a plain leaning. It is a transparent heuristic, not a real detector. It uses no trained model, and even purpose-built detectors are unreliable, so the result is a hint and never proof of authorship.
It gives a quick read, with the maths shown, on whether writing looks uniform and repetitive (which can read as machine-generated) or varied (which tends to read as human). You see the score, the leaning and the two raw figures behind them.
It does not decide who or what wrote the text. It is a writing-rhythm indicator built from surface statistics, so you can see the signals instead of trusting a black-box verdict.
As you paste text, it splits it into sentences, measures the spread of sentence lengths and the ratio of unique words, and prints a leaning that updates live.
The whole calculation is an open formula running in your browser. There is no language model, remote API or training data, which is why the page calls it a rough estimate.
Sentence variation (burstiness) = standard deviation of words per sentence
Vocabulary variety = unique words ÷ total words
Score = 100 − (variation × 6) − (variety × 40), clamped to 0–100
Verdict: above 60 leans AI · 40 to 60 mixed · 40 or below leans human'The tool is fast. It is private. It is free.' → variation 0.5, vocabulary 70%, score 69 → leans AI-like (uniform short sentences)The 6 and 40 weights are arbitrary tuning, not a validated model, which is one more reason to treat the number as a loose hint.
Your text is analysed locally in your browser. Nothing you paste is sent to a server, stored or logged.
There is no account and no history. Refresh the page and the analysis is gone, so you can paste unpublished drafts and confidential work.
| Method | Sentence-length variation plus vocabulary diversity |
|---|---|
| Signals used | Burstiness (standard deviation) and type-token ratio |
| Score range | 0 to 100, with a three-band verdict |
| Minimum input | At least two sentences |
| Model | None. It is an open formula, not a trained classifier |
| Processing | 100% in-browser, live |
This is not a definitive detector. No tool can be. It does not read meaning or use a trained language model. It measures only two surface signals, how much your sentence lengths vary and how often words repeat, so it is a rough indicator and never proof.
Even purpose-built, commercial detectors are unreliable. OpenAI retired its own classifier in 2023 after it caught barely a quarter of AI text and flagged human writing as fake. Treat any AI-detection result, including this one, as a weak hint.
It can be wrong both ways. Terse, repetitive human writing, such as a checklist or a beginner's essay, can score AI-like, while edited or deliberately varied AI text can score human-like. Short inputs are especially unstable.
The vocabulary measure naturally falls as text gets longer, so a long human passage can drift toward an AI-like score for reasons that have nothing to do with how it was written.
Never use this to accuse a student, writer or employee. Detectors are known to be biased against non-native English speakers, and a false accusation does real harm.
Paste your own draft to see whether it reads as uniform or repetitive, then add sentence-length variety. That is useful editing feedback regardless of how the text was written.
Show students why varied sentence rhythm matters, using a transparent formula instead of a black-box score.
If everything scores uniform, try breaking up same-length sentences. That is good prose advice anyway.
See, in the open, how crude the underlying signals are and why AI-detection results should not be trusted as evidence.
There is no reliable way to prove AI authorship from text alone. Even leading commercial detectors get it wrong. Use this only as a writing-rhythm hint, never as evidence.
Checked against this tool's code: it scores text only from sentence-length variation and the unique-to-total word ratio, with no trained model behind it. That is why the page treats every result as a hint, not a verdict.
Picks the 1 to 8 sentences that matter most. Rule-based.
Trims your opening sentence to about 155 characters.
Fixed plain-English edits that tighten wordy text.
A fill-in blog post skeleton built around your topic.
Lists the most frequent words and repeated phrases in a text.
A rough positive, negative or neutral read from a word list.
It's rule-based and runs in your browser, so your text stays with you. No account needed.