Instant AI detection.
Our models find “human signal” in any text with a keen focus on preventing false detection. Uses range from academic integrity to resume authenticity.
On a regular basis, we publish when we are wrong with supported reasoning. A flagged document should infer additional follow-up. It is not proof that anyone cheated.
Paste 50 to 1,200 words of English. One check a day without an account. Our detector scores the text in memory and de-identifies any personal information for additional training. Your browser keeps the text for 7 days so you can sign in and get a verification link.
Roles
Flagged result?
Choose the tab for your role.
Hand in verification links
Check your work then share the provided link with your teacher or editor. The link provides critical information that can support your authenticity.
If your work is flagged AI
Keep your drafts and version history. Do not run your writing through a rewriting tool or humanizer to change the result. Humanizers are notorious for making up false information that can be detected by our models.
Results
Three results and why
Every check ends with one of three results.
Verified human
The mark
Verified human: the detector read the text as human. The result includes the text and a date stamp.
Likely AI
With the sentences
The check could not confirm the text as written by a person and shows the sentences that read as AI. We include the reason and the likely AI percentage.
Not scored
Under 50 words
Not enough text to read. Our model will not return a reading under 50 words, so there is no result to give.
A flag should be reason to look closer and not verified proof of cheating.
Limits to detection
There are cases where a result deserves less weight.
Humanizer tools
Our model called 20 of 176 sealed humanized AI papers human. These can be detected with our deep search that finds false facts often created by humanizers.
The test is on the model cardAuthorship
AI detection models, including ours, do not confirm authorship. For example, another person other than the author could have written the work.
Short texts
Under 50 words will return “Not scored.” 50 to 149 words results in a weaker reading.
English writing only
Our model only reads English academic and narrative writing.
Writing before 2022
Our model was trained on human-written documents written before 2022.
No outside check yet
We invite outside parties to check our results. Email support@olive.is.
Not a plagiarism check (yet)
Our model does not check plagiarism. Coming soon.
Education
Free for teachers
Sign up with a school or university email and check up to 50 papers a month free, with 15 deep checks a month, in the web app or MCP server.
Higher education
Departments, writing centers and integrity offices: get authenticity links students can hand in.
High school
School and district wide at affordable rates.
Middle school
Teacher accounts only. No accounts for students under 13.
A department or campus
Seats and a shared deep-check allowance.
Verified human: our model read the text as human. Our detector does not catch every rewrite written by a humanizer. However, our deep check can detect false facts often created by humanizers that hallucinate.
One check a day needs no account, up to 1,200 words. A free account gets 2,000 words a day and a link for each check. With a school email, a teacher checks up to 50 papers a month free.
No. A flag is a reason to look closer. “Not verified” means the check could not confirm the text. We provide a likely AI flag by sentence and paragraph.
Our standard detector flagged 0 of 600 human-written papers in a release test (95% upper bound 0.50%). Flagged 1 of 549 held-out human papers as Mixed (upper bound 0.86%). Flagged 5 of 9,305 real student essays, all as Mixed, none as AI. Missed 0 of 481 wholly machine-written held-out papers (upper bound 0.62%).
Yes. Our model called 20 of 176 sealed humanized AI papers human. In a newer test, after one humanizer tool rewrote machine-written texts, it called 231 of 255 long papers and 122 of 136 essays human. While a humanizer tool can get text past our model, they do so at a significant cost with a tell-tale sign: false facts like wrong dates, incorrect citations and wrong information. Our deep check can find inaccuracies.
The standard detector flagged 0 of 1,195 essays by English learners. The students detector, shown as a note, flags 4 of the same 1,195.
Our detector scores the text in memory and de-identifies the writing for additional training. A verification link keeps an additional keyed fingerprint of the text.
A deep check looks for references outside of Crossref or OpenAlex, unusual stock phrases, pairs of numbers, names or dates worth a second look, and up to 8 dates and numbers checked against Wikipedia.
Not enough text to read. Readings under 150 words are also weaker.
No. Our model reads English academic and narrative writing. It is not built or tested for other languages.
Not yet. We invite outside parties to contact us at support@olive.is to access our datasets for independent testing.
Start with writing you know.
Paste something you wrote and see the result. With a school email, a teacher checks up to 50 papers a month free.