AI detector
Paste an essay, article or email to see which sentences read as machine-written. We check for ChatGPT, Claude, Gemini and other models, and show our working one sentence at a time.
Try a sample
How to read your score
Every scan returns a human score from 0 to 100. It is our confidence that a person wrote the text, not the share of the text a person wrote.
A 72 means we are fairly confident a person wrote it. It does not mean 28% of it came from a model.
- 0–34Likely AI-generatedMost sentences match patterns we see in model output. Review the highlights, then talk with the writer before acting on it.
- 35–59Mix of AI and humanParts read as human and parts as AI. Common when a draft has been edited, paraphrased or extended with an assistant.
- 60–100Likely human-writtenThe writing has the uneven rhythm and specific detail people produce. A few highlights are normal and rarely mean much alone.
Scores on texts under 150 words swing more. Lists, code and heavily quoted passages are also harder to judge, so lean on the sentence highlights there.
What the highlights mean
The score sums up the whole text. The highlights show where it came from, one sentence at a time, so you can see exactly which lines raised it.
Our spring volunteer drive raised more than we expected, mostly thanks to Dana's lemon bars. In today's fast-paced world, community engagement serves as a cornerstone of a thriving neighborhood. (AI probability 92%) Moreover, it fosters meaningful connections that transcend generational boundaries. (AI probability 86%) Planning ahead will help us reach more families and keep the momentum going into the fall. (AI probability 54%) Next year we want to start earlier and maybe skip the raffle, which nobody understood. Thanks again to everyone who showed up with a folding table.
- Red: likely AI
- The sentence closely matches how language models write: stock transitions, safe word choices and an even rhythm.
- Yellow: mixed signals
- Some AI-like patterns, but not enough to call. Often a human sentence that was polished or paraphrased by a tool.
- No highlight: reads as human
- The sentence looks like something a person wrote. In most honest writing, the majority of sentences stay plain.
Models we check for
The detector learns from writing produced by the models people actually use, and we retrain it as new versions ship. It tells you whether text is likely machine-written, not which product wrote it.
- ChatGPT
- Claude
- Gemini
- Llama
- Mistral
- DeepSeek
- Grok
- Copilot
- and their newer versions
Languages
Scan text in 14 languages. We detect the language automatically and show it on your report.
- English
- Spanish
- French
- German
- Portuguese
- Italian
- Dutch
- Polish
- Turkish
- Arabic
- Japanese
- Korean
- Chinese
- Indonesian
Questions about the detector
Something else on your mind? Visit the help center or read how we test accuracy.
On our benchmark set, Winston AI labels text correctly 99.4% of the time, with a false positive rate of 0.6%. Accuracy drops on very short texts, lists and heavy paraphrasing, so read the sentence highlights rather than the score alone.
No, and we don't claim to. The detector looks for patterns that models such as ChatGPT, Claude and Gemini share. The result tells you how likely it is that text was machine-written, not which product wrote it.
Light edits usually don't move the score much. Heavy rewriting by a person can lower it, which is fair: at that point a person did much of the writing. Paraphrasing tools leave tells of their own, and we test against the common ones.
Free scans are analyzed and then discarded. When you're signed in, reports are saved to your account so you can come back to them, and you can delete any report at any time.
At least 40 words. Below that there isn't enough writing to judge and results swing too much to be useful. For the most reliable score, scan 150 words or more.
No. A score is one piece of evidence. Look at the highlighted sentences, compare them with the writer's earlier work and have a conversation before reaching any decision.