About Textlume

Writing tools that tell you where they fall short

Textlume checks and edits text, and says plainly how sure it is. Our detector publishes its own error rates, and our rewriting tools leave your facts alone.

Textlume is our set of writing tools. It has an AI detector, a humanizer, a paraphraser, a grammar checker, a summarizer, two counters and a plagiarism checker that is still on the way. We built it for people who have to judge a text fairly, or fix one without changing what it says.

We earn money from paid plans and one-off word packs, paid by card. The grammar checker and both counters are free with no limit. The detector, humanizer, paraphraser and summarizer each give one free run a day, with no card needed. We do not use your texts to train models.

Who we build for

A student checking a history paper the night before it is due. A teacher who wants a second look before raising a concern. An editor going through a freelancer's draft. An engineer from Brazil making sure a report for a US client reads naturally. Each of them needs more than a label. They need to see why a tool said what it said, and what it could not tell.

They also need to know where their text goes. If you are not signed in, nothing you paste is stored. Signed-in accounts keep a history for up to a year, and you can switch it off or delete it at any time. The word and character counters do their counting in your browser.

How the detector reads a text

The detector runs on our own detection model. It scores each sentence in the context of the passage around it, then reads that score against a line fitted separately for each language, on human and machine writing that includes non-native writers. It does not hunt for watermarks or banned words.

Every result is a score with a range and a High, Medium or Low confidence rating, plus a mark on each sentence. A text under 50 words gets no reading, and one under 100 words gets Low confidence. Human writing can still be flagged, more often in some languages than others. We publish how often.

How the rewriting tools work

The humanizer and the paraphraser rewrite with a large language model through an AI service that may not keep your text or train on it. Both keep the names and numbers in your draft as they were. The humanizer has four tabs, Standard, Academic, Professional and Casual, and on paid plans it writes three candidates per run.

We aim for text that reads like a careful person wrote it and still says what your draft said. We make no promise about what any detector, ours included, will say about the result. A humanizer that promised a pass would give you good reason to doubt our detector.

What we hold ourselves to

Publish the error rate
The accuracy page shows how often the detector flagged human writing in each language, with the number of texts behind every rate. Weak results stay on the page.
A reading, not a verdict
A high reading is a reason to look closer and talk to the writer. It is never a finding about a person.
Count non-native writers
Writing in a second language is tested as its own group. A language is calibrated only if no more than 1 of those texts in 20 crosses the line.
Keep the facts intact
Rewrites keep the names, numbers and quotes you wrote, so a smoother sentence never changes what you claimed.
Your text stays yours
We do not use your texts to train models. Signed-out text is not stored, and you can clear your history whenever you like.

Our figures, with their limits

These are English results from our latest held-out test, on texts the detector never saw in training. The accuracy page has every language with the count behind each rate, and our weak spots.

See the accuracy results

The people behind the guides

Four editors and engineers write and check what we publish. Jakob Wendt builds and tests the detector and checks each technical claim against the code. Carmen Echeverría edits the academic writing guides, Ece Karaman leads the guides for students and teachers, and Malcolm Thorne edits the guides on writing for work.