Free AI Text Detector

What this detector misses

Most free detectors publish an accuracy number and stop. Here is ours, and then here is the part nobody publishes.

In September 2026 we generated fresh AI text from five current models, in the same two genres as our human control text so that genre alone could not separate the classes. At the threshold this tool ships with, it catches half of AI text while wrongly flagging about 1% of human text.

That half is not spread evenly:

Frontier-model writing is effectively invisible to this detector. Not harder to catch — zero of twelve Claude Opus 5 passages and zero of eight Qwen passages scored above the threshold. If someone used a current top-tier model, this tool will almost certainly tell you the text is clean, and it will be wrong.

That is not a flaw specific to this detector. It is the state of small open detection models in 2026, and it is why a clean result here means very little.

Why the threshold is 0.97, not the model's default of 0.5

A detection model outputs a probability, and someone has to choose the line at which that probability becomes an accusation. Nearly every free detector built on an open model keeps the stock line of 0.5. We measured what that costs, using text that cannot be AI-generated because it was published before ChatGPT existed: 155 paragraphs from pre-2022 English Wikipedia revisions and 183 paragraphs from open-access academic papers published 2015–2021.

We give up 25 points of detection to get false positives near one percent, because a false accusation costs a real person far more than a missed detection costs us.

How badly an easy benchmark lies

The first model we evaluated looked outstanding: AUC 0.9985, 98.8% detection. Then we changed the test set three times, and each change was simply a move toward honesty.

Its "human" class was pre-1940 public-domain literature. Replacing Jane Austen with pre-2022 Wikipedia raised its false-positive rate from 4.6% to 20.6%. Its "AI" class was mid-2026 GPT-4o text in a different genre from the human text. Replacing it with current-model text in matched genres dropped its AUC from 0.9935 to 0.6711 — barely better than a coin flip.

Same code, same metric. Only the test set changed. Whenever you see a detector advertising 98% or 99%, this is the most likely explanation.

Why there is no sentence-by-sentence highlighting

Most free detectors highlight individual sentences. We measured where this model fails, and it fails on short text. Across 306 passages of public-domain literature, zero of the 296 passages over 150 words were flagged — but 10% of those under 100 words were. A 70-word passage of Dickens scores 0.974, above the threshold. Jane Austen, Charlotte Brontë and the Gettysburg Address all score higher at that length than they do at full length.

Sentence-level highlighting produces confident-looking colour on exactly the input length where the model is least trustworthy. So this tool scores windows of 150–350 words and refuses anything shorter.

How to read your result

No AI signal detected (score under 0.90)

This means very little. The detector misses half of all AI text and all frontier-model text, so a clean result is not evidence of human authorship.

Uncertain (0.90–0.97)

The least reliable range. Treat it as no result — not as a soft accusation.

Likely AI-generated (0.97 and above)

The only band that carries real information. Even here, about 1% of the human text we tested landed in it.

If you have been flagged

No detector output is proof of anything, and no honest one claims otherwise. If a piece of your own writing is flagged, the useful evidence is your process, not a second opinion from another detector: version history in your editor, drafts, notes, and timestamps. Ask whoever flagged you what false-positive rate their tool was measured at, and on what text. Most cannot answer.

Limitations

• A score is a probabilistic signal, not proof of authorship. Do not use it alone for a disciplinary, academic, hiring or moderation decision.
• English only.
• Raw, unedited learner writing is untested. Our non-native-English evidence comes from published academic papers, which are professionally copy-edited. The corpora that would test unedited learner English are all research-licence-only and cannot be used in a tool anyone sells.
• It misses half of all AI text, and 100% of Claude Opus 5 and Qwen output. A clean result is close to meaningless.
• 56 AI passages and 338 human passages is a small evaluation. The ~1% false-positive threshold rests on roughly three passages and is correspondingly uncertain.
• Two genres were tested, encyclopedic and academic. Student essays and business copy were not.
• Paraphrased or "humanized" AI text was not tested at all, and is built to defeat exactly this kind of model.
• Each window is truncated at 512 tokens.

Frequently asked questions

Is my text uploaded anywhere?

No. The model file is downloaded to your browser and the analysis runs on your own device. You can disconnect from the network after the model loads and it still works.

Why is there a 150-word minimum?

Because we measured that below roughly that length the model produces confident false positives on text known to be human — 10% of sub-100-word literary passages, versus zero above 150 words. A tool that accepts 20 words is not being generous, it is hiding its error rate.

Can it detect AI text that has been rewritten or paraphrased?

Often not. Paraphrasing tools exist specifically to remove the statistical regularities this model keys on. Treat a clean result on suspected paraphrased text as uninformative.

Which model is this?

An int8 ONNX build of tmr-ai-text-detector (MIT), 126 MB, run with transformers.js. It downloads once and your browser caches it. The threshold and every number on this page are ours.

Is it the same detector as the paid product?

No. This is a small open model chosen because it fits in a browser, and the numbers above are its numbers, measured by us. It is not the production detector and says nothing about it.