Students
Run a history paper or a college essay before you hand it in. Keep your drafts, since they are your best backup.
Your result
A figure for the whole text and a mark on every sentence
Sample detector result
Clear communication plays a vital role in building strong and effective teams. By fostering open dialogue, organizations can greatly improve collaboration. On my team, the best fix was a ten-minute call every Tuesday at nine. Regular updates help people know what everyone else is working on.
6 tools
Read the result, then work on the draft
AI detectorMarks each sentence as AI-like or human-like and gives the whole text a range and a confidence rating.OpenAI humanizerRewrites stiff AI drafts into plainer prose and checks that numbers, names and quotations come through.OpenGrammar checkerFinds spelling, grammar and punctuation errors and lets you fix them one by one or all at once.OpenParaphrasing toolSays the same thing in new words, in the mode you pick, from fluent to formal.OpenSummarizerCuts a long text to a short, medium or long summary, or pulls out its key sentences.OpenWord counterCounts words, characters, sentences and paragraphs as you type, with reading and speaking time.Open
Our test results
We publish how often we flag human writing
- How often human texts were marked as AI, language by language
- Separate figures for writers working in a second language
- Sample sizes and the test date next to every figure
Languages: 13
Each language tested on writing in that language
- English
- Français
- Español
- Português
- Deutsch
- Italiano
- Русский
- हिन्दी
- العربية
- Türkçe
- Polski
- Nederlands
- Bahasa Indonesia
Pricing
Start free and pay only when you need more words
Free
0 USD
One run per day on the detector, humanizer, paraphraser and summarizer, with up to 250 words in a detector check. Grammar checks and both counters have no daily cap.
Sign upBasic
6.99 USD / month
Billed yearly. 50,000 AI detector words, 10,000 humanizer words and 10,000 paraphraser words a month.
Choose BasicPlus
9.99 USD / month
Billed yearly. 150,000 detector words a month, with 25,000 for the humanizer and 25,000 for the paraphraser.
Choose PlusPro
12.99 USD / month
Billed yearly. 300,000 detector words a month, with 50,000 for the humanizer and 50,000 for the paraphraser, plus the API, custom style and downloadable detector reports.
Choose ProHow sure we are
Every reading comes with a confidence rating
High
The evidence agrees
A longer text in a language we have tested, with sentences that point the same way. The headline says it straight, for example "Reads as human-written".
Medium
A lean, with the evidence shown
Some sentences are less clear-cut. The headline softens to "Likely" and the sentence marks show you which lines pull the figure up or down.
Low
A weak signal, and why
The text is short, our two models disagree, or the sentences split both ways. You get the reason in one line, so a single figure never has to carry the decision.
Quick start
How to use the AI detector
The AI detector gives its clearest reading on a full passage, like an essay section or a blog draft.
Paste or upload your text
Paste a few paragraphs or upload a Word file or PDF. The language is read from the text, and you can change it under Language.
Run the check
Press Check for AI. One plain sentence at the top says what the reading means.
Start with the marks
Open Most AI-like sentences to jump to the strongest signals, then read each one in context.
What it does
What an AI detector can and can't tell you
An AI detector estimates how much a passage reads like a language model's writing. Textlume scores each sentence in the context of the whole text, against human and AI writing in the same language and of a similar length.
It doesn't prove who wrote anything. It doesn't search for sources, and it can't name the model behind a text. Some human writing will read as AI, and some AI writing will slip through. That is why every result carries a range and a rating instead of a yes or no.
Method
How the reading works
Textlume runs your text through our own detection model on our own GPU servers, and a second, independent model reads it too. When the two land far apart, the range widens and the rating drops.
Thirteen languages are calibrated, each on its own. Within a language, texts under 100 words, texts of 100 to 249 words and longer texts each get their own curve, because length changes how much the model can see.
A 2023 Stanford study found that popular detectors flagged essays by non-native English writers far more often than essays by native speakers. Textlume tests on learner writing and publishes the result. Where a language and length band still flags human writing too often, the rating stays Low and the card says the limit is ours.
Example
What the marks look like
As we move forward, it is essential to ensure alignment across all key stakeholders. This approach will strengthen collaboration and drive meaningful results for the entire team. Quick heads-up: the Denver client pushed Thursday's call to after lunch. Clear priorities usually make the week go smoother. If someone can grab the slides from the shared drive, I owe you a coffee.
Your result
How to read your result
The headline, such as Reads as AI-written or Likely human-written, is scaled to how sure the reading is. Below it sit the figure, its range and the rating. The figure describes the whole passage. It isn't the share of AI sentences.
The range shows how far the reading could move. Short text or two models that disagree will widen it, and a wide range means a Low rating.
Each sentence is marked AI, human or unclear, and a deeper shade is a stronger reading. You only see Mixed when marks fall on both sides. When the two models disagree sharply, no sentence gets a side at all.
Accuracy
How we measure AI detector accuracy
A single accuracy number hides the two errors that matter: AI writing missed and human writing flagged by mistake. Our accuracy page shows both for each language, with sample sizes, from a test on texts the detector never saw in training or tuning.
In that test of October 12, 2026, the detector flagged 0 of 150 native English human texts and 3 of 600 English texts by non-native writers. It is weaker at catching machine text in some languages than others, French most of all, and the accuracy page shows that too. Our rule for a language: no more than 1 human text in 20, and no more than 1 non-native writer in 20, may land on the AI-written side of the line.
Who uses it
Who checks their writing with it
Teachers
Use a reading to start a conversation. Ask how the student built the argument and what changed between drafts. A student who wrote the paper can usually walk you through it.
Editors and content teams
Check a freelancer's blog draft before it goes live, alongside your fact checks. If a section reads as AI, ask the writer about it before you edit.
Two different tools
AI detector or plagiarism checker?
An AI text detector reads how the writing is built. Our plagiarism checker will look for the same wording on pages across the open web once it opens for checks.
| AI detector | Plagiarism checker | |
|---|---|---|
| What it compares | Your writing against patterns in human and AI text | Your wording against text published online |
| What a flag means | The passage reads like AI to our models | Some wording matches a source |
| What it can't show | Who wrote it, which model, or whether it was copied | Whether the matching wording came from AI |
| What to do next | Weigh the reading against your drafts | Open the source and check quotes and citations |
Flagged?
If your own writing gets flagged
Flags on human writing happen, and one reading is never the whole story. Answer a flag with evidence.
Reread the marked sentences
See which sentences carry AI marks. Polished, formulaic lines can draw marks even when you wrote every word.
Check a longer sample
If you checked one paragraph, run the whole section. More text narrows the range.
Pull your writing records
Collect your outline and the document's version history. A timeline of edits shows how the piece grew.
Walk someone through it
Explain how you wrote it and share the range and rating, so the uncertainty is out in the open.
Your text
What happens to your text
Signed out, nothing kept
When you're signed out, your text is read for the check and nothing is saved.
History you control
When you're signed in, your checks stay in your history for up to one year. Delete entries whenever you like or turn history off.
Fix the language next
Then run the draft through the Textlume grammar checker for spelling and punctuation. It's free.
Regular checks
For people who check every week
Batch checks
Paid plans check several texts in one go: up to 10 on Basic, 20 on Plus and 50 on Pro. Export the results as a CSV file.
Printable reports
On Pro, a check becomes a printable report that you can save as a PDF. It stays only in your browser, for 24 hours.
API access
Pro includes the API, so you can send texts for checking from your own scripts and apps.
Is there a free AI detector?
Yes. Textlume gives you one free check a day without an account, each up to 250 words. A free check gets the full reading: the range, the confidence rating and a mark on each sentence. The pricing page shows what paid plans add, such as more words and batch checks.
Do AI detectors actually work?
As a signal, yes. As proof, no. Every detector misses some AI text and flags some human text. In our October 2026 test, Textlume caught 68 of 75 English machine texts and flagged none of 150 native English human texts. None in 150 is not a promise of never: for a sample that size, the honest ceiling on false flags is about 2.5 percent. That is why each reading comes with a range and a confidence rating. The accuracy page has the counts for every language, weak spots included.
Can it check text from ChatGPT, Gemini or Claude?
Yes. Textlume reads the text the same way whichever assistant produced it, so it works as a ChatGPT detector and for the others too. It looks at the writing, not where it came from, so it won't tell you which assistant wrote a passage.
Can teachers use an AI detector on student essays?
Yes, as one input to a conversation and never as the final word. Textlume shows which sentences read as AI and how confident the reading is. Put that next to the student's drafts and version history, and be extra careful with a Low rating or with a writer still learning English.
Which languages can it check?
Textlume is calibrated for 13 languages: English, Spanish, German, French, Italian, Portuguese, Dutch, Polish, Russian, Turkish, Arabic, Hindi and Indonesian. It reads the language from your text, and you can set it yourself. A text in any other language gets no calibrated figure. The accuracy page has the results for each language.
How many words do I need to paste?
At least 50 words, and more is better. Anything shorter is turned away. From 50 to 99 words, Textlume gives a reading but holds the rating at Low and tells you why. A full essay section or a complete email gives a clearer picture.
Is the text I paste saved?
Not if you're signed out. The text is read for the check and nothing is kept. Signed in, each check stays in your history for up to one year, and you can delete entries or turn history off. Checks sent through the API aren't kept either.
Why would my own writing be flagged as AI?
Plain, polished, predictable prose can look like model output, and detectors have a known habit of flagging non-native English writers, which Textlume measures and publishes. Editing tools matter too. A grammar checker that fixes spelling and commas leaves your sentences as you built them, while a tool that rewrites whole sentences hands the wording to a model, and that wording is what a detector reads. If a flag seems wrong, check a longer sample and keep your drafts ready.
Can it show which parts of a text are AI-written?
It marks each sentence as AI, human or unclear, but it can't prove who wrote any one of them. Textlume says Mixed only when marks land on both sides. The overall figure reads the whole text and is not a count of AI sentences.
Does it catch paraphrased or humanized AI text?
Sometimes. Reworded AI writing can still read as AI, but we publish no separate figure for humanized text, so don't read a human result on a reworked draft as a clean bill. For English text, Textlume also adds a note when its second model sees signs of AI editing or a humanizer.
Can it tell which AI model wrote a text?
No. Textlume reads whether a text looks machine-written, not which model wrote it. Our test set covers 9 current chat models, plus continuations and paraphrases. One known weak spot: a text where a model carries on from a human-written opening is caught far less often.