AI Detector: Check if Text Was Written by AI

Perplexity: how predictable

Language models pick likely next words, so their text tends to be predictable. A detector with access to a model scores how surprised that model is by each word. Low surprise points toward AI.

Burstiness: how varied

People mix short and long sentences and change rhythm as they think. AI output is often more even. Low variation in sentence length and structure is a second signal.

Classifier models

Many commercial detectors train a model on large sets of human and AI examples and let it learn the difference. They work best on text like their training data and less well on new models or unusual writing.

The “estimated AI-generated likelihood” summarizes how strongly the whole text matches AI patterns. It is a probability-style estimate, not a measurement of how much was typed by a model.

“Likely AI”, “Mixed / uncertain” and “Likely human” split the text into three groups. A large mixed share is normal for edited drafts and text written with AI help.

Red, amber and green underlines show which sentences drove the result. Hover a sentence to see its own estimate, then ask why it reads that way.

Non-native English writers

Writers with a smaller working vocabulary use more common words and simpler structures, which look “predictable”. In a Stanford study, seven detectors labeled 61% of 91 TOEFL essays by non-native writers as AI-generated.

Formulaic writing

Academic abstracts, legal clauses, technical documentation, cover-letter templates and five-paragraph essays follow fixed patterns. Clear, conventional human writing can score as AI simply because it is conventional.

Mixed and edited text

When people and AI both touch a document, sentences near the joins are hard to call. Turnitin reports a sentence-level false-positive rate of around 4%, with many of those sentences sitting right next to AI-written text.

Short texts and new models

A few sentences don't give a detector enough to measure. And each new generation of models writes more like people, so detectors trained on older output fall behind.

You wrote it yourself

• Keep notes, outlines and earlier drafts. Version history in Google Docs or Word is strong evidence of your process.
• Look at the highlighted sentences. They are usually generic or formulaic lines, not your whole text.
• If you're questioned, explain how you wrote it and ask that the result is weighed with other evidence.

You used AI help

• Check the rules for your class, publication or employer, and disclose AI help where it's required.
• Rewrite the draft in your own voice: add your own examples, opinions and specifics, and cut filler phrases.
• Fact-check every claim. AI drafts can sound confident and still be wrong.