Know where words came from.

Paste a draft — yours or anyone’s. Nota tells you whether it looks like AI wrote it, finds where its text came from, and spots hidden characters. Every score comes with its measured error rate. Built for students, schools and organisations that want evidence, not a verdict.

250words minimum. Under that, Nota abstains and tells you why instead of faking a number. Pasted text is discarded unless you choose to keep it.
Samples:
0 words · minimum 250
  • Measured false-positive rate on every result
  • Abstains under 250 words
  • Never “written by AI”
  • Pasted text discarded by default
  • A reason to look closer, not a verdict
  • Model version printed on every score

What an honest AI detector looks like.

An honest read on whether text looks like a machine wrote it. Not a verdict. Not a compliance stamp. Three things follow from that.

01

A number with its error rate

Every scan prints the probability, the threshold, the band and the false-positive rate measured for the exact model that scored it — returned by the API, not typed into a marketing page.

02

Abstention instead of guessing

Under 250 words Nota declines to score and says why. It does not name a generator, does not highlight sentences, and does not split “human / mixed / AI”. It scores documents whole.

03

Evidence beside the score, never inside it

Writing measurements against a human envelope, a source check that aligns matched passages, and a character check that reports hidden marks and look-alike letters. Each labelled for what it is — inferred, measured, matched or observed — and never averaged into the Nota number.

Who it is for

Built for people who read a lot of writing — and want to be fair about it.

The same model, the same caveats, whatever you pay. What changes between plans is volume, batches, reports and API access.

Students

Check your own draft before you hand it in. See how machine-like it reads under one named model — and how often that model is wrong about human writers, printed right beside the score.

Free plan: 2,000 Nota words a day, no card. · More →
Schools

A number with its error rate, not an accusation. Nota abstains on short text, names no generator and highlights no sentences — so a result opens a conversation instead of closing one.

No published ESL false-positive rate yet. We say so rather than sell one. · More →
Organisations

Editorial desks, publishers, hiring and comms teams: batch files, signed Nota Records you can file, and an API that returns the caveats with the probability.

Meter Nota words. Text is processed in memory and discarded unless you keep it in a record.
On every result

A report, not a traffic light.

  1. Wording, not a verdict. “Consistent with machine generation under this model.” Never “written by AI”, never “X% AI”.
  2. Band first, then AI-likeness. The band leads (low, moderate, elevated, high). The 0–100 reading is resemblance under this model — never a percentage of authorship.
  3. Measured false-positive rate. How often this checkpoint flagged held-out human documents of 250+ words.
  4. Model version and word count. So a result can be reproduced and understood later.
  5. The caveat. Always present, never collapsed.
Nota · model version printed here
Consistent with human writing under this model
Moderate bandp = probability
reads humanreads machine-made

Overlap: careful human writing and light machine polish both live here.

Measured FPRfrom the APIheld-out humans, 250+ words
Operating pointfootnote1% FPR
Wordscountedmin 250

Layout illustration. Values appear only when Nota has scored a real text.

The model

Nota, in one paragraph.

Nota is a small language model with a classification head, trained on pre-2022-style human writing and AI mirrors of the same kinds of documents, and examined on frozen test sets. It reports how machine-like a passage looks under that model, with the false-positive rate we measured on held-out humans. It cannot name the generator, and it will not score anything under 250 words.

Document-level

A 4-billion-parameter language model with a LoRA adapter and a linear classification head. Whole documents; no sentence maps.

Trained on mirrors

Human writing is pre-2022-style. Machine text mirrors the same kinds of documents, from many model families, and the exams are frozen — so the published rate is the rate.

Measured, then shipped

Every checkpoint carries the false-positive rate measured on held-out humans of 250+ words. That number rides along in the API.

Read the method
Claims we do not make

If a detector promises these, be careful.

Sentences you will not find on this site, in our API responses or in our sales decks.

“This was written by AI.”We say “consistent with machine generation under this model”. The difference is the whole product.
“Detects ChatGPT, Claude or Gemini.”Nota is trained on the output of many generators, but it does not name them. It scores how machine-like the text reads, and prints the error rate beside the score.
“EU AI Act compliant.”No certification scheme exists, and using a detector does not transfer anyone’s obligations.
“99.9% accurate.”A single flattering number hides the false-positive rate. Ours is measured for the shipped model and returned with every result.
“Catch AI cheating.”A score is a reason to look closer, not an accusation. We do not sell it as one.

Free to try. Pay for volume, reports and API.

Nota words are what we meter. The evidence panel comes with every scan on every plan.

Free
$0no card
  • Nota: 2,000 words a day
  • Text of 250 words or more
  • Evidence panel on every scan
Start free
Professional
$59per month · $45 billed annually
  • Nota: 1,200,000 words a month
  • API access with Bearer keys
  • Larger batches (up to 100 files)
Choose Professional
Developer
Creditsprepaid, metered by the 100 words
  • Nota words, rounded up to 100
  • Bearer keys, usage by endpoint and day
  • Same JSON as the app — FPR and caveat included
Read the API docs

Questions people ask

Straight answers. The longer versions are on the method page.

Is a high score proof that AI wrote this?

No. It means the text is consistent with machine generation under this model, at a threshold with a measured false-positive rate. Careful, formal human writing can score high; paraphrased machine text can score low. Treat it as a reason to look closer, never as an accusation.

Which AI models does Nota detect?

Nota is trained on the output of many different generators, so it reports how machine-like a text reads — full stop. What it does not do is name the culprit: it cannot tell you whether ChatGPT, Claude or Gemini produced something, and we do not claim it can.

Why does it refuse to score short text?

Below 250 words there is not enough text to score responsibly. Rather than print a confident-looking number, Nota abstains and says why — paste a longer passage of continuous prose.

Where does the false-positive rate come from?

It is measured for the shipped checkpoint on held-out human writing of 250 words or more, and returned by the API with every result. We do not print a marketing number that could drift from the model actually scoring your text.

Do you store what I paste?

Not unless you ask us to. Text is processed in memory to produce a result and discarded. Signed-in scans keep a signed record — the SHA-256 of the text and the numbers — and the text itself only if you switch on “keep the document in my record” for that scan.

Paste 250 words. Read the number beside the score.

No account for the first look. Pasted text is discarded by default.

Try Nota free