01
Perplexity
Measures how predictable the text is. Machine generated text usually scores lower because it follows the patterns the model was trained on.
Academic integrity toolingVersion 2.0
AI text detection, with evidence.
Detect AI-generated writing with an ensemble of four independent analysis methods — perplexity, burstiness, transformer classification and stylometry.
No account needed for a quick checkSentence-level evidenceFull text is never stored
- —
- Texts analysed
- Platform activity
- —
- Balanced accuracy
- Grouped cross-validation
- —
- Labelled samples
- Coai benchmark · English
- 04
- Analysis methods
- Independent signals
Example output
Calibrated ensembleAI generated
Sentence-level evidence, not a black box.
91.8
AI probability
Perplexity88
Burstiness74
Transformer96
Stylometry82
Sentence-level breakdown12
Human writtenAI generated
Analysis desk
Examine a manuscript
Paste a passage, upload a document or fetch a public page. Minimum 50 characters.
Full text is never stored
ManuscriptVersion 2.0
Try an example:
0 words · 0 characters
At least 50 characters required (50) more needed
Full text is never stored
Method
Four independent signals
Each analyser scores the text on its own. The final verdict is their weighted combination, which is far more reliable than any single method.
02
Burstiness
Analyses the variation in sentence length. Human writing is far more varied, while AI text tends towards uniformity.
03
Transformer
Uses a RoBERTa network trained specifically to recognise content written by GPT-family models.
04
Stylometry
Examines vocabulary richness, word patterns and writing style markers that separate human from machine authors.
No account needed for a quick checkSentence-level evidenceFull text is never stored
How it works
From pasted text to a reasoned verdict
No black box: every score is broken down into independent signals you can inspect.
Paste or upload
Paste an article, upload a .txt, .docx or .pdf file, or fetch a public URL. Formatting is stripped and the text is split into sentences.
Four independent checks
A transformer classifier, language-model perplexity, sentence rhythm and writing style each rate the text. The scores are blended with language-aware weights.
Read the evidence
You get an overall probability, a per-sentence map and the contribution of every signal — plus an interpretation you can quote.
FAQ
Questions people ask before trusting a score
Honest answers about what this detector can and cannot do.
How accurate is the detector?
It is a decision aid, not proof. Accuracy is highest for English academic prose and lower for short, edited or translated text. The interface shows how confident the combination of signals is; treat the percentage as evidence, not as a verdict.
Which languages are supported?
The interface is available in English, Uzbek and Russian. Detection works best for English. For Uzbek and Russian the style-based signals carry more weight because the language models were trained mostly on English.
Is my text stored?
Only a short preview and the score are saved to your account history so you can revisit them. The full text is not stored. You can delete individual records or your entire history at any time.
Why can the same text get a different score?
Each of the four signals measures something different, and they do not always agree. When they disagree the reported confidence drops. Editing, translation and very short passages all move the score.
Can I use the result in a report?
Yes — download the JSON report or copy the summary. For formal or disciplinary decisions, always combine the result with human review and additional evidence.
Academic integrity tooling
Check a text before you publish it
Paste an article, a book chapter, an essay or a letter. The evidence behind the score is one click away.