ChatGPT detector that shows which sentences gave it away
Check any text for the fingerprints of ChatGPT. You get an overall likelihood, sentence-by-sentence highlighting, and a plain list of the habits that made the text read as machine-written.
0 / 1,500 words
Your report will appear here
What ChatGPT writing tends to look like
Left to its defaults, ChatGPT writes in a recognisable house style. Paragraphs come out at similar lengths. Sentences keep an even pace. Points arrive in threes. Transitions like “Moreover” and “In conclusion” do the work of connecting ideas, and a handful of phrases (“plays a crucial role”, “it's important to note”, “in today's fast-paced world”) appear far more often than they do in human writing.
None of these habits proves anything alone. People write this way too, especially in school essays and corporate copy. The detection model looks at the statistics of the whole text; the pattern list shows you the visible part of what it is reacting to.
Edited and prompted ChatGPT text is harder
Ask ChatGPT to write casually, give it your own notes to work from, or rewrite half its sentences yourself, and the result drifts toward the human end of the scale. Every detector struggles with this, ours included. Expect those texts to land in the Uncertain band, and read the sentence-level highlights to see which parts still carry the model's voice.
If you used ChatGPT and want to know how it reads
Plenty of people check their own drafts. If you brainstormed or outlined with ChatGPT and wrote the rest yourself, the highlights show where the model's phrasing survived. Rewriting those sentences in your own words, with your own examples, usually improves the writing as well as the score.
How the AI detector works
An AI detector is a classifier that estimates the probability that text was written by a language model rather than a person. It works by measuring statistical patterns in the writing, not by tracing where the text came from.
Give it something to read
Paste text, drop in a PDF or Word file, or enter the address of a web page. Documents are opened in your browser, so the file itself never leaves your device.
The text is scored
A detection model trained on human writing and on output from AI systems rates the whole text and each sentence for how closely it matches machine writing.
You see the evidence
The report marks every sentence, lists the writing patterns found, and tells you how much confidence the result deserves.
How accurate is AI detection, honestly?
AI detection is accurate enough to raise a question, but not accurate enough to prove an answer. The best detectors catch most unedited AI text with few false alarms, but accuracy drops sharply when text is paraphrased, lightly edited, or written by a non-native English speaker.
Key takeaways
- Accuracy claims of 99% come from the vendors’ own benchmarks.
- One 2026 test of five detectors put the top two at 72% and 74% overall accuracy.
- False positives fall hardest on non-native English writers.
- A score is a probability estimate, not evidence of misconduct.
Reading the score
| Score | Verdict |
|---|---|
| 0–14 | Human-writtenLittle or nothing resembles AI writing. |
| 15–34 | Likely human-writtenMostly reads as a person, with a few AI-like passages. |
| 35–69 | UncertainMixed signals. Not evidence either way. |
| 70–89 | Likely AI-generatedMostly reads as AI, possibly edited by a person. |
| 90–100 | AI-generatedStrong, consistent AI signals throughout. |
A score in the Uncertain range is reported as Mixed when the sentences back that up, with at least 25% of the words in AI-like sentences and at least 25% in human-like ones.
Supported languages: English, French, Spanish, Portuguese, German, Dutch, Polish, Romanian, Italian and Simplified Chinese.
ChatGPT Detector: common questions
Does it work on the latest ChatGPT models?
The detection model is trained on output from the GPT family among others and is retrained as new models come out. Newer models write more naturally than older ones, so detection is never perfect; the confidence level on each result tells you how much weight to give it.
Can it tell ChatGPT apart from Claude or Gemini?
No. It estimates whether a language model wrote the text, not which one. The major assistants share most of the habits a detector picks up on.
Will it catch ChatGPT text that was run through a paraphraser?
Sometimes. Paraphrasing and “humanizer” tools reduce detection rates across the board. A low score on paraphrased text is weak evidence that a person wrote it.
How much text do I need?
At least 50 words to get a result, and 250 or more for one you can lean on. Very short answers do not contain enough signal for any detector.