Reverse Turing Test

Ensemble verdict

—

Per-model breakdown

Prediction flow

Human AI Ribbon colour is the prediction each model contributed.

Feature analysis

Feature	Your text	Human avg	AI avg	Closer to

Or try a labelled sample

Every sample is a real row from the held-out RAID test split, so the badge is verified ground truth and no model saw it during training. The two research abstracts share a subject on purpose — one human, one generated — since telling those apart cannot be done on topic alone.

How it works

Three independent classifiers each read the text, and a fourth model — a stacked meta-classifier fitted on held-out data — decides how much to trust each one.

Transformer — a fine-tuned DistilBERT reading the text as language.
Stylometric — logistic regression over sentence rhythm, vocabulary richness, and punctuation.
Visual CNN — renders the text as a heatmap image and classifies the texture of its structure.

All three run in your browser — the text you paste is never sent anywhere. See the method page for measured accuracy.