AI Content Detector
Paste any text to get a heuristic score of AI-style writing patterns. It is not a calibrated probability and does not identify specific models. Free — 5 checks per day with a Google account.
How to detect AI-generated text
AI language models like ChatGPT, Claude, and Gemini produce text with characteristic patterns: unusually consistent sentence length, high-frequency "filler" transitions ("Moreover", "It is worth noting"), heavier em-dash use, and repeated sentence openers. This detector scores those patterns directly with a fixed heuristic — no text is sent to an external AI model to generate the result, and the score is not a calibrated statistical probability (see below).
For best results, submit at least 150–200 words. Heavily edited or paraphrased AI text will score lower on the AI-likelihood meter, because the human editing obscures the original model's stylistic fingerprint.
This score is a heuristic, not a calibrated probability. It is not benchmarked against a labeled validation dataset, and no confusion matrix, false-positive rate, or subgroup analysis (e.g. by language or writing style) is published for it. Independent research has also found that AI-text detectors in general can systematically misclassify non-native-English writing and can be defeated by simple prompting (arxiv.org/abs/2304.02819). Treat the result as one input among several, not a verdict.
Frequently Asked Questions
It scores writing patterns known to differ between human and AI-generated prose — sentence-length variation, AI stock phrasing, em-dash frequency, repeated sentence openers, and contraction use — against a fixed heuristic, not a trained statistical model. No text is sent to an external AI model.
It has not been benchmarked against a labeled dataset, so no accuracy, false-positive rate, or per-language figure can be honestly stated for it. Treat the score as a heuristic signal, not a validated probability. Independent research shows AI-text detectors in general can misclassify non-native-English writing and can be defeated by simple prompting (see arxiv.org/abs/2304.02819) — the same caution applies here. Heavily edited AI text, paraphrasing, or very short passages make any detector, including this one, less reliable.
Login lets us enforce fair-use limits (5 checks per day) so the tool stays free. We store only your Google ID, email, and name — we never read or post anything on your behalf.
It does not identify specific models. It scores general writing patterns that are common in LLM output, so it cannot tell you whether text came from ChatGPT, Claude, Gemini or any other particular model, and it can flag human writing that happens to share those patterns.
No. We only log metadata (character count, result label, probability). The text you paste is never saved to our servers.