What Textumi is for
- Checking your own draft before you send it: the detector highlights the passages that sound generic or machine-made and deserve another look.
- Rewriting a text you are entitled to edit, such as your emails, articles, website copy or an AI-assisted first draft, so that it reads more naturally and keeps its meaning.
- Getting one more signal about a text you received, as a reason to read it more carefully and never as a verdict on its author.
You can work wherever it suits you: in the browser, in the macOS app or right in a conversation with your own AI assistant through the MCP connector. We keep improving the detector and the Humanizer and adding tools that make working with them easier.
Either way, a person reads the result and decides what to do with it. The tools speed up that work. They do not replace it, and they take no responsibility off the person who acts on the result.
No promise about any AI detector
The Humanizer makes a text read more naturally and keeps its meaning. Before you see a result, Textumi checks the facts, names, numbers, exact names and logic against the original, and a version that fails those checks is discarded.
It does not make a text “undetectable”, and Textumi does not sell it as a way to pass, bypass or beat a detector. Detectors are built differently and often disagree, so a text one of them accepts can be flagged by the next, whoever wrote it.
If a result with a particular detector decides something important for you, a rewriting tool is the wrong place to look for safety. The dependable course is to follow the rules you are being checked against and to be able to show how the text was written.
A detector result is a signal, not proof
The Textumi detector is our own development: we train and tune its models ourselves, and they run on our own servers. It names the likely author with one of five labels, from “Human-written” to “AI-generated”, depending on how clearly the text shows patterns typical of machine writing, and it highlights the passages that weighed most.
That is a statement about wording, not about who sat at the keyboard, and it can be wrong in both directions. Keep the known limits in mind:
- Short texts are harder to judge: the detector needs at least 5 words, and under 50 words it shows “Too little text to judge” instead of a label.
- The English and Ukrainian detectors are the most thoroughly tested. The detectors for other languages were trained and tested on a narrower set of texts, prose from Wikimedia projects.
- Outside that genre some of them miss many AI texts, among them French, Vietnamese, Turkish, Indonesian and Dutch.
- Stanford researchers reported in 2023 that AI detectors flag the writing of non-native English speakers more often than that of native speakers.
- OpenAI withdrew its own AI text classifier in 2023 and gave its low accuracy as the reason.
Why detectors flag human writingHow the Textumi detector is tested
Do not decide about a person on a detector result alone
A grade, a disciplinary case, a hiring decision or a contract cannot rest on a detector result alone, whether it comes from us or anyone else.
A “Mostly AI” or “AI-generated” result is a reason to look further, and people have to do that: talk to the author, ask how the text was produced, look at drafts, notes and version history, and compare the text with the author's earlier work.
If you are the one being judged by a detector, the same evidence is your best answer. Keep your drafts and sources. A document history that shows a text growing over several days is more convincing than any tool.
Academic integrity: your institution's rules come first
Schools, universities and journals set their own rules on AI help, and they differ. Some forbid it in assessed work, some allow it for language editing, and many ask you to disclose it. Textumi does not know which rule applies to you and cannot change it.
Read the policy for your course or journal, and ask if anything is unclear. Disclose AI help wherever disclosure is required. The Humanizer does not change how a text was first created, and it does not turn undisclosed help into your own unaided work.
Using Textumi to hide what a rule requires you to disclose is not what we built it for.
Other people's texts and privacy
Textumi stores submitted texts and results encrypted. The core processing runs on our own servers; for a rewrite, the text is also sent to the external AI model provider named in the privacy policy. On the Individual and Professional plans a request stays in your private History for 30 days unless you delete it sooner.
So before you submit something another person wrote, make sure you are allowed to share it. For a student's essay, an applicant's letter or a colleague's report, check what your institution or employer allows, and remove names and personal details the check does not need.
My style is stricter still: it accepts only texts you wrote yourself, and you confirm this before the analysis begins.
Who should not rely on Textumi
- Anyone who needs evidence of authorship for a formal accusation or a legal dispute, because a detector result is not such evidence.
- Anyone judging very short texts, such as single answers, headlines or chat messages, because Textumi gives no label under 50 words, and short texts in general give a less reliable result.
- Anyone working mostly in languages other than English and Ukrainian, or in genres far from plain prose, where the detector misses more AI text.
- Anyone ready to send a rewritten contract, medical instruction or financial figures without reading them. Textumi checks a rewrite in several ways, but some of those checks are done by a model, and a model can miss an error.
Common questions
Does Textumi guarantee that a text will pass AI detection?
No. Textumi makes no promise about how any detector, its own included, will judge a text after a rewrite. The Humanizer makes text read more naturally and keeps its meaning. Detectors differ from one another and change over time, and they sometimes flag writing no machine touched. Treat any guarantee of an undetectable text as a warning sign.
Can a detector result prove that someone used AI?
No. A result describes patterns in the wording and says nothing certain about who wrote the text. Human writing can get “Mostly AI”, especially formal prose and writing in a second language, and edited machine text can get “Mostly human”. Use a result as a reason to read more closely and ask questions, never as the finding itself.
Should teachers use an AI detector to grade students?
Not as the ground for a grade or an accusation. A detector can point to a passage worth a conversation. The decision should rest on that conversation and on the student's drafts, notes and earlier work. Before you submit student work to any online service, find out what your institution's data protection rules allow.
Is it acceptable to humanize AI text for school or university?
Only within your institution's rules. If an assignment forbids AI help, rewriting does not make it allowed. If AI help is allowed with disclosure, disclose it, and mention that you used a rewriting tool if the policy asks for that level of detail. If the rules are unclear, ask your teacher before you hand the work in.
What should I do if my own writing is flagged as AI?
Stick to the facts of how you wrote it. Show your drafts, notes, sources and the document's version history, and offer to walk someone through the text. Point out that detector results are signals with known error, especially for short texts and writing in a second language. Rewriting the text afterward to lower the result proves nothing and can look worse.
Can I check a text that someone else wrote?
Yes, if you have the right to share it. Textumi stores texts encrypted, and if your plan includes History, you can delete a saved request at any time. Remove personal details the check does not need, follow your organization's rules for other people's documents, and remember that the result is not proof about the author.
Which detector languages are the most reliable?
The English and Ukrainian detectors are the most thoroughly tested. The detectors for other languages were trained and tested on a narrower set of texts, prose from Wikimedia projects and text from two AI models. Some of them, including French, Vietnamese, Turkish, Indonesian and Dutch, miss many AI texts in other genres. In every language a result is a signal, not proof of authorship.