AI Text Detector

An AI text detector that shows its working — and its limits

This AI text detector looks for the statistical habits of machine-written text and tells you exactly what it found. It never says “99% AI”. You get one of three cautious answers — more like typical AI text, mixed, or more like typical human text — with an uncertainty band, the signals behind it, and a sentence-by-sentence map.

Everything runs on your device. Writing statistics are instant and need no download. For a stronger signal, an open language model (SmolLM2 from Hugging Face, Apache-2.0) can measure how predictable your text is, in your browser, after you agree to a one-time download. Your text is not uploaded.

Use it to understand a text, not to judge a person: a result here should start a conversation at most — never end one.

How to check whether a text reads like AI writing

1. Paste at least 80 words into Text to check. Longer is better: the band narrows with length.
2. Read the result. Writing statistics alone are weak evidence, so most texts come out as “Mixed”.
3. Open the signals table to see each measurement next to the human and AI texts in our test set.
4. For a stronger check, choose one model (133 MB) or two (about 700 MB) and press Run the model check. The page asks before downloading.
5. Look at the sentence map and the reasons, then read the limits box before drawing any conclusion.

AI Text Detector features

How AI detection works, and why it fails

Most detectors use one idea: language models write the words they themselves find most likely, so their text has low perplexity. GPTZero’s early explanation (1 March 2023) paired perplexity with burstiness — how much that predictability varies through a text; it says it moved to a deep-learning model in autumn 2023. DetectGPT (Mitchell et al., ICML 2023) compares a text with slightly reworded versions of it. Binoculars (Hans et al., ICML 2024) divides the perplexity from one model by a cross-perplexity between two related models, which cancels much of the effect of topic.

Each breaks in predictable ways. OpenAI withdrew its own AI classifier on 20 July 2023 “due to its low rate of accuracy” — it had caught 26% of AI-written text while flagging 9% of human text. Liang et al. (Patterns, July 2023) found seven detectors labelled on average 61% of TOEFL essays by non-native English writers as AI-written, and 97.8% were flagged by at least one. Watermarks such as Google DeepMind’s SynthID-Text (Nature, 23 October 2024) are more robust, but only the vendor can check them with its key — no outside detector, including this one, can see them.

How this page was calibrated — and how often it was wrong

The thresholds come from 48 English texts of 97–273 words. The 24 human texts are public domain: ten Project Gutenberg excerpts (1839–1918) and fourteen US federal pages and notices written before November 2022. The 24 AI texts were written by Claude (12, varied styles) and by IBM’s Granite 4.0 350M (12). Half set the thresholds; half were held back. No human text was labelled “AI” — but many AI texts were missed, especially the large model’s.

Worked examples (writing statistics only)

• “AI, stock phrases” sample: 174 words, 10 sentences averaging 17.4 words with a spread of only 26%, and 5 stock phrases (“In today’s fast-paced world”, “plays a crucial role”…). Score 0.96, band 0.36 to 1 → More like typical AI text.
• “Human, health page (2019)”: no stock phrases and a 41% spread. Score −0.55 → Mixed. With the model it moves toward the AI side (0.48) because public-health instructions are very predictable — a false positive in the making.
• “AI, casual”: a text we asked an AI to write casually. Short and long sentences, contractions, no stock phrases. Score −1 → More like typical human text. That is the false negative every detector has.

If you are a teacher, an editor or an employer

Treat any detector score as weak, unverifiable evidence. Talk to the writer, ask about drafts and sources, and compare with work you know they wrote — a falsely accused student can’t prove a negative. If you want help making AI-assisted drafts read naturally — openly, where it is allowed — the AI Text Humanizer finds stock phrases and flat rhythm. For reading level, try the Readability Checker.