Free AI Detector
- Free Forever
- No Sign Up Required
- Unlimited Usage
- Your Text Never Leaves Your Browser
How it works
Most AI detectors are a web form in front of somebody else’s server: you paste your text, it is uploaded, and you hope it is deleted. This one inverts that. The model is sent to your browser instead of your writing being sent to a server, which means the tool can be free, works offline once loaded, and never puts unpublished work in someone else’s logs.
- 1
The model comes to you
When the page opens, an open-source AI-detection model starts downloading in the background and is cached by your browser. You can start writing straight away — the scan runs the moment it is ready, and on later visits it is already there.
- 2
Your text is split into passages
The model can read about 500 tokens at once — roughly 360 words — so the text is divided at sentence boundaries into passages of about that size. Each passage is scored with its full context in view, which is where the model is most accurate.
- 3
Each passage is scored on its own
Every passage is read in full and produces one probability. The document score is the word-weighted average of those passage scores, so the headline number and the breakdown underneath it always describe the same set of measurements.
- 4
Then each passage is broken into sentence groups
Every passage is split again into short runs of sentences — a sentence stands alone when it is long enough to score, and joins its neighbours when it is not, so nothing is ever set aside as too short. Each group is read on its own to see where inside the passage the score comes from, then the groups are levelled onto their passage’s score. That is what the heat map colours, and it means the sentence percentages average back to the headline number exactly.
How accurate is it?
The detector is a RoBERTa-base classifier published as onnx-community/tmr-ai-text-detector-ONNX under an MIT licence. On the RAID benchmark it reports 99.28% AUROC and a 95.79% true-positive rate at a 5% false-positive rate.
Those figures describe performance on a benchmark, not a guarantee about your document. Be sceptical of the result when:
- the text is not in English — the model was trained on English alone, so its score means nothing on any other language;
- the text is short — under about sixty words, treat any score as noise;
- the writer’s first language is not English, where false positives are a well-documented and serious problem;
- the writing is technical, legal, templated or otherwise naturally uniform;
- the text was AI-drafted and then substantially rewritten, or vice versa;
- the text is a translation, or contains long quotations, references or code;
- the text came from a generator the model was never trained against — including models released after it.
Signs of AI drafting
These are the habits that give generated prose away to a reader.
Even sentence rhythm
Style tab: Sentence-length variationSentence after sentence lands at roughly the same length, usually somewhere between fifteen and twenty-five words. Human drafts are lumpier: a long clause-heavy sentence, then a short one.
What to do: Break one long sentence in two. Let a four-word sentence stand on its own. Read the paragraph aloud and notice where you run out of breath.
Stock transitions
Style tab: Stock AI phrasingFurthermore, Moreover, Additionally, It is important to note that, In conclusion. Signposting on every joint, as though the reader could not follow the argument without one.
What to do: Cut them and reread. If the link between two paragraphs stops making sense without a signpost, the problem is the argument, not the missing word.
The em-dash and semicolon habit
Style tab: Em-dash, semicolon & colon rateEm dashes several times a paragraph, semicolons joining clauses that a full stop would separate, colons introducing lists that did not need introducing.
What to do: Keep the ones doing real work and turn the rest into full stops. Varying the punctuation usually means varying the sentence structure too, which is the actual improvement.
Repeated phrasing
Style tab: Repeated three-word phrasesThe same construction recurring every few paragraphs, or a favourite noun phrase used four times where a pronoun would do.
What to do: Say it once, properly. Repetition is worth keeping only when it is doing rhetorical work you intended.
Abstract nouns doing the work
Style tab: Vocabulary varietySentences built on utilisation, implementation, optimisation and engagement, where nobody in particular does anything in particular.
What to do: Name the actor and the verb. Who did what to whom? "The implementation of the policy resulted in improved outcomes" becomes "After the policy changed, waiting times fell by a third."
No specifics
No names, no dates, no numbers, no first-hand detail. Everything is true in general and about nothing in particular. This is the strongest sign on this list, and the hardest for a classifier to measure.
What to do: One concrete particular per paragraph: a figure, a date, a person, something you saw. Specifics are also what makes writing worth reading.
Hedged, symmetrical conclusions
Every section ends by weighing both sides and committing to neither. "While there are benefits, there are also challenges."
What to do: Decide what you think and say it. If you genuinely do not know, say that instead — it is a real position, and it reads as one.
Uniform structure
Paragraphs of near-identical length, sections of near-identical shape, or a piece of prose that has quietly turned into a listicle with three bullets under every heading.
What to do: Let the structure follow the argument. A point that needs two sentences should get two sentences, even if its neighbour needed nine.
Frictionless prose
Nothing is qualified, nothing is reconsidered, nothing is awkward. No asides, no second thoughts, no sentence that was clearly hard to write.
What to do: Keep your digressions and your caveats. The texture of someone thinking is not a defect to be polished out.
Limitations and responsible use
No AI detector can prove how a piece of text was written, and this one makes no such claim. It produces a probability, and probabilities are wrong sometimes — in both directions.
Please do not use this score as the sole basis for an academic-misconduct, hiring or disciplinary decision. Detector output has already led to students being wrongly accused, and the burden falls hardest on those whose writing is least like the model’s idea of “natural” English. Use it as one signal among several: look at drafts and version history, ask the writer about their process, and weigh the context. A high score is the start of a conversation, not the end of one.
Frequently asked questions
Is this AI detector really free?
Yes. There is no account, no trial, no word limit behind a paywall and no card required. The detector runs on your own device, so there is no per-scan cost for us to pass on to you.
Does my text get uploaded anywhere?
No. The detection model is downloaded to your browser once and cached there. After that, every scan happens on your device in a background thread. Your text is never transmitted to Evidano or to any third party. You can verify this by opening your browser developer tools, switching to the Network tab, and watching that no request carries your text — or by disconnecting from the internet after the model has loaded and scanning anyway.
Why is there a one-time download?
Because the detection happens locally, the model itself has to reach your device. It is about 126 MB, or 250 MB if your browser supports GPU acceleration and can use the higher-precision build. Your browser caches it, so you only pay that cost once and the page then works offline.
How accurate is it?
The underlying model reports 99.28% AUROC on the RAID benchmark and a 95.79% true-positive rate at a 5% false-positive rate. Those figures were measured against the text generators that existed when it was trained, and they are not a guarantee about your document. We publish our own measurement against newer generators in the accuracy section above; when that measurement does not show a clear separation between human and AI writing, the tool shows the model output and withholds a verdict rather than asserting one.
Can it produce false positives?
Yes, and you should assume it will sometimes. False positives are measurably more common for writers whose first language is not English, and for technical, formulaic, templated or academic prose that is naturally regular. A high score is a reason to ask a question, never a finding of fact.
Can I use this to accuse a student of cheating?
Please do not. No AI detector — this one included — produces evidence that can carry an academic-misconduct or disciplinary decision on its own. Use the result as one input among many: talk to the writer, look at drafts and version history, and consider the context. Treating a percentage as proof produces unjust outcomes, disproportionately for non-native English speakers.
What do the human, mixed and AI percentages mean?
Every passage gets its own score. Passages at or below the lower threshold are grouped as varied, those at or above the upper threshold as formulaic, and everything between as mixed. The three percentages are measured in words, so a long passage counts for more than a short one. You can move both thresholds in the settings. These are groupings of a style score, not claims about who wrote each passage.
Can it score individual sentences?
Yes, on demand, but as a ranking rather than a verdict. Open a passage and it scores each sentence on its own. We measured this against passages with known AI and human sentences spliced together: per-sentence scores rank the AI-written sentences well (ROC-AUC about 0.88, against 0.61 for giving every sentence its passage score), which is what the ranking uses. What they do not do is give a trustworthy percentage — the same measurement shows the model inflates the score of short spans, so a lone sentence often reads as more AI-like than it is. That is why sentences are shown as an order, most to least AI-like within a passage, with the shortest ones set aside, and not as standalone percentages.
Why does it not just tell me whether text is AI-written?
Because we measured it and it cannot do that reliably. We ran the model over human writing of certain provenance and over text from current models. The two groups overlapped almost completely: a news report from 2013 scored 64%, while an essay written by a current model minutes earlier scored 2%. What consistently moved the score was register, not authorship — formulaic, impersonal, uniformly paced prose scores high whoever wrote it. Rather than dress that up as a verdict, we report it as what it is: a reading of how formulaic your writing looks. The measurement is published in the accuracy section on this page, and if a future model earns the stronger claim the tool will make it.
Which file types can I upload?
Plain text, Markdown, PDF, Word (.docx) and CSV. Each one is parsed inside your browser — the file is never uploaded. Scanned PDFs that contain only images have no selectable text, so nothing can be extracted from them locally.
Does it work in every browser?
It needs a modern browser with WebAssembly, which covers current versions of Chrome, Edge, Firefox and Safari. Browsers with WebGPU run it considerably faster. On iPhones and iPads the download is opt-in rather than automatic, because those devices have tighter memory limits.
Does it work on languages other than English?
No, and we measured it rather than guessing. We ran the detector over public-domain human prose published between 1605 and 1939 in ten languages, and over chatbot-written essays in the same ten. In English it separated the two by 86 percentage points. In the other nine it failed, in one of two ways. On French, Spanish, Italian, Portuguese and German it rated the chatbot text as human — AI-written French scored 2%, lower than Victor Hugo. On Russian, Arabic, Chinese and Japanese it rated the human text as AI — Natsume Sōseki writing in 1905 scored 75%. Both errors are worse than no answer. The tool checks the script and the common words of whatever you paste and puts a warning above the score when the text does not look like English. Word counts, the passage breakdown and the writing-style signals still work in any language; only the AI percentage should be ignored.
What happens when I edit the text?
By default the tool re-scans about a second after you stop typing, and only when the edit changed something meaningful — a stray keystroke will not trigger it. Only the passages whose text actually changed are re-run, so a re-scan is usually near-instant. The Auto-scan control above the editor lets you turn it off, switch to a fixed timer instead of waiting for a pause, change how long that pause is, set how many words must change first, and decide whether pasting or clicking away should trigger a scan. Ctrl+Enter (Cmd+Enter on a Mac) always scans on demand.
Can I compare two versions of my text?
Yes. Every scan you run is kept in the History tab for as long as the page is open, and the compare view lines the passages of any two of them up against each other. It ranks the edits by how much of the overall move each one accounts for, so you can see which rewrite pushed the score up or down rather than guessing. Passages are large, so an edit is attributed to the whole passage it falls in rather than to the words you changed: the ranking explains the direction and the relative size of a change, not the headline figure exactly.
Is my scan history saved, and where?
Scans of sixty words or more are saved in this browser, on this computer, using IndexedDB — the same on-device storage your settings use. Nothing is uploaded; there is no server to upload it to. The store keeps the twenty-five most recent documents for up to thirty days, one row per document rather than one per scan, so editing does not fill it with near-identical copies. You can pin a scan so it is never evicted, delete any single scan, delete all of them at once, or switch saving off entirely from the History tab.
What are the telltale signs of AI writing?
Even sentence rhythm, stock transitions like "Furthermore" and "In conclusion", a heavy em-dash and semicolon habit, repeated phrasing, abstract nouns where a person and a verb should be, and an absence of specifics — no names, dates, numbers or first-hand detail. None of these is proof on its own; plenty of careful human writing is regular, and plenty of AI output is not. The Style tab measures the four of these that can be counted, and the guide below the tool explains what each one means and how to revise for it.
What model does this use?
onnx-community/tmr-ai-text-detector-ONNX, an MIT-licensed RoBERTa-base classifier converted to ONNX and run through Transformers.js. It was trained on English text only, so its score means nothing on other languages.