AI Writing Signals Analyzer
See the patterns that read as AI
Your text is processed locally by this tool — it isn't sent to Wibble. Free, no account.
Works best with 100+ words. Sentence splitting is heuristic, so abbreviations like Dr. may over-split.
How it works
Paste at least a hundred words and the analyzer breaks your text into sentences, words, and paragraphs, then computes the measurable traits that make prose read as machine-written: how uniform your sentence lengths are (the coefficient of variation), how often you open sentences with formal transitions, which phrases repeat, your lexical diversity (a moving-average type-token ratio), and standard readability scores. Each number comes with a plain-language reading. It deliberately never outputs an 'AI probability', percentage, or verdict.
Example
Reading the sentence-uniformity number
Two passages can share the same average sentence length but read very differently. If almost every sentence runs 18-20 words, the coefficient of variation drops below 0.3 and the analyzer labels it 'very uniform' — the metered cadence readers and detectors associate with unedited AI output. Human drafts usually mix short and long sentences, pushing the value past 0.5 into 'varied'. Those 0.3 and 0.5 cutoffs are Wibble editing rules of thumb, not established norms — we haven't published corpus data behind the exact numbers.
Limitations
- It is not an AI detector and produces no score, percentage, or verdict — high uniformity or heavy transitions are correlations, not proof, and plenty of edited human writing shows the same patterns.
- The numbers need roughly 100+ words to mean anything; on shorter passages they're computed but noisy, and the tool says so.
- Sentence splitting is a simple heuristic that over-splits on abbreviations like 'Dr.' or 'e.g.', which nudges the sentence-length and readability figures.
- The transition and vocabulary lists are fixed English sets, so another language, or a field where 'crucial' and 'however' are normal, will read as higher without meaning much.
- Readability uses the Flesch formulas, which estimate syllables heuristically and were designed for English prose, not code, lists, or tables.
Frequently asked questions
Is this an AI detector?
No, and it avoids being one by design. It reports observable measurements — sentence uniformity, transition density, lexical diversity, readability — and explains what each means, but it never outputs an 'AI probability' or verdict. Tools that give a single score are guessing; this shows you the underlying signals instead.
Does my text get uploaded?
No. All the analysis runs locally in your browser. Your text is never sent to Wibble or any external service, stored, or logged. Once the page has loaded you could disconnect from the internet and the analyzer would still run.
What counts as a 'good' coefficient of variation?
There's no pass/fail line, and the cutoffs we use are Wibble editing rules of thumb rather than established norms — we haven't published corpus data behind the exact numbers. As a heuristic, human prose is usually bursty: the CV of sentence lengths often sits above 0.5, which the tool labels 'varied', while values under 0.3 ('very uniform') mean nearly every sentence is the same length — the even rhythm associated with unedited AI text.
Why does short text show a warning?
Below about 100 words there aren't enough sentences and word types for the statistics to be stable, so a small edit can swing the numbers a lot. The tool still computes everything, but flags that the results say little until you paste a longer, representative sample.
What is MATTR and why use it over a plain type-token ratio?
MATTR is a moving-average type-token ratio measured over a 50-word window. A plain type-token ratio always falls as text gets longer, so long and short samples aren't comparable. MATTR averages diversity across fixed windows, giving a length-robust view of how varied your vocabulary really is.
Wibble rewrites these patterns into natural prose — 300 words free.
Try Wibble free