Nobody has published a large, rigorous, independent accuracy study of QuillBot’s AI detector that we could find. QuillBot says it has “high accuracy” and also that “no approach is 100% reliable”. The few outside tests are small and disagree in places, so treat its result as a rough signal about which passages look machine-written, not as proof of who wrote a text.
Key takeaways
- QuillBot’s own pages claim accuracy but publish no error rates, test sets or conditions we could find.
- Independent checks are small: one used four texts, another sixteen. They are anecdotes with useful details, not benchmarks.
- The detector reports more than yes or no. It separates AI-generated, AI-refined and human text, which is useful and also harder to get right.
- Features such as free limits and languages change. Check QuillBot’s current page before relying on any number here.
- A flag is a reason to look closer, never a verdict.
What does QuillBot say its AI detector does?
QuillBot’s AI detector article (read 2026-10-13) describes four result labels: “Generated by AI”, “Generated and refined by AI”, “Written by a human and refined by AI” and “Written by a human”. Elsewhere it refers to AI-generated, AI-refined and human-written content. The free report gives percentages for AI-generated and AI-refined content, but the page does not say how those percentages are calculated.
The same page says scans under 1,200 words are free and need no sign-up, and that the detector “currently supports six languages: English, French, Spanish, German, Dutch, and Portuguese.” Bulk upload is a Premium feature. Third-party reviews have listed different limits and language sets at different times, which suggests the product changes. Confirm the current terms on QuillBot’s own site.
QuillBot’s FAQ on accuracy says the tool “effectively differentiates between text produced by AI and by humans” and suggests combining it with manual review of tone, style and consistency. It gives no figures. We could not open QuillBot’s main detector page, which returned an access error to our fetch, so we rely on these two pages.
Has anyone tested it independently?
Only informally. Two public tests give some detail, and neither should be read as a benchmark.
SlashGear, December 2024. A writer ran four texts through several free detectors: a human-written LinkedIn article, a ChatGPT article on the same subject, a human-written book preface and a ChatGPT-written fiction passage. Both human texts scored 100% human. The AI factual text was flagged as AI but with 34% judged human-written, and the AI fiction with 7% judged human-written. Four texts cannot support a general claim.
IntoTheMinds, August 2025. A market-research firm tested free tools on 16 texts, half AI and half human, in French and English. QuillBot flagged 0% of the eight human texts as AI. On the AI texts its share flagged ranged from 47% to 100%, averaging 82.5%, with the weakest scores on text from a newer ChatGPT model. The firm’s own business is market research, and the sample is small and partly written by the author.
Put together, these tests point the same direction at best: few false alarms on the clean human texts used, and misses on some AI text. That is a hint, not a measurement. We saw other accuracy figures online, such as 78% or 80%, but could not trace them to a test with stated conditions, so we do not repeat them. GPTZero, a competitor, has also published a review that says QuillBot may lean toward flagging human text. That is a vendor’s opinion of a rival, so weigh it accordingly.
What did the two tests find, side by side?
| Test | Texts | False alarms on human text | Detection of AI text | Main limit |
|---|---|---|---|---|
| SlashGear, December 2024 | 4 (2 human, 2 AI) | None; both human texts scored 100% human | AI texts flagged, with 34% and 7% judged human-written | Four texts |
| IntoTheMinds, August 2025 | 16 (8 human, 8 AI), French and English | 0% of the eight human texts flagged as AI | 47% to 100% flagged, averaging 82.5% | Small sample, partly written by the author |
Neither test controlled for text length, subject or editing, and neither reports a confidence interval. The table shows what each reported and nothing more.
How does QuillBot compare with other free detectors?
On what you can check yourself, the free tools differ mostly in limits, not in published accuracy. In our comparison of free AI detectors, checked on 2026-10-13, QuillBot’s free scan was 1,200 words, the same ceiling Scribbr listed, while GPTZero listed a 10,000-character box and Copyleaks up to 25,000 characters. None of those pages publishes error rates by dataset in the way that Pangram’s model card does.
That means we cannot rank QuillBot against its rivals on accuracy, and a page that does so without a stated test is guessing. If you want a second view, our guides on GPTZero, Grammarly and Copyleaks set out what each vendor claims and what outside studies found.
Sources for this guide were read on 2026-10-13. We have not re-tested the detector since, and QuillBot changes its limits and labels, so check its current pages.
Why is the “AI-refined” category harder than it looks?
Telling fully AI text from fully human text is already imperfect, as how accurate AI detectors are explains. A middle label asks for more: whether a human wrote the text and an AI tool polished it, or a model wrote it and a person edited it.
Those cases overlap with ordinary writing help. Grammar fixes, a paraphrasing pass and a heavy human edit of a draft can all leave similar traces. The detector has no access to a draft history, so the label is an inference from the surface of the text. For definitions of these categories, see AI-assisted vs AI-generated writing.
How should you read a QuillBot result?
| What you see | What it can reasonably mean | What it does not show |
|---|---|---|
| Mostly “human” | Little of the text resembled the AI patterns the model learned | That a person wrote it. AI text that was edited or paraphrased can pass |
| Mostly “AI-generated” | Many passages resemble model output | That the author cheated. Formal, generic or translated human writing can resemble it |
| “AI-refined” share | Some passages look polished by a tool | Which tool, how much, or whether it was allowed |
| A mix across the text | Different passages scored differently | That the mix is accurate. Check the highlighted sentences yourself |
Check these things before acting on any result:
- Length. Short samples give noisy results. Rerunning a longer extract is more informative than trusting one score.
- Language. Outside the languages QuillBot lists, expect less dependable results.
- The highlighted passages. Read them. Are they generic, conventional or heavily edited? Are they the parts a person would have had the most reason to polish?
- The writer. Non-native English, formal genres and translated text are known trouble spots; see why your writing may show as AI.
- The meaning of the percentage. It is a share of text, not a probability about the author. What an AI detection percentage means covers the distinction.
Does a second detector settle it?
Not by itself. Two tools agreeing is more informative than one, but detectors can share blind spots, and they can disagree for ordinary reasons: different training data, different thresholds, different labels. If results differ, that is a sign the text sits in a gray zone, not that one tool is secretly right.
Treat a second tool as one more weak signal. Stronger evidence comes from the writing process: drafts, version history, notes and sources, and a conversation with the writer. Our guide on telling if text is AI-generated lists the checks that carry more weight than any score.
What should you do with the number?
If you are checking your own writing, use the highlights to spot passages that sound generic, then revise for clarity and accuracy and keep your drafts. Do not tune text to move a score; a lower number is not a better essay, and ai detection false positives shows how often honest work gets flagged.
If you are judging someone else’s work, do not penalize on a detector result alone. QuillBot itself advises checking institutional guidelines and not treating any approach as fully reliable. What to do when a detector flags student writing gives a fair process.
Whatever the tool, ask what a mistake would cost and who bears it. For the general case, see can AI detectors be wrong.
Frequently asked questions
Is QuillBot’s AI detector free? QuillBot’s own page says scans under 1,200 words are free and need no sign-up, with bulk upload reserved for Premium. Terms have varied across third-party reviews, so check the current page.
Does QuillBot’s detector work in other languages? QuillBot’s page lists English, French, Spanish, German, Dutch and Portuguese. It says accuracy is high in each, but publishes no per-language figures we could find. Treat non-English results as less certain.
Can the QuillBot detector be wrong? Yes. QuillBot says no approach is 100% reliable, and every detector can flag human writing or miss AI text. Short, formal, translated and edited text is where mistakes are most likely.
Should I trust a QuillBot result as evidence of cheating? No. It reports how text resembles patterns, not who wrote it. Use it to decide what to look at next, then rely on drafts, sources and a conversation.
Want a second look at your own text?
If you already have the text, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score. We do not publish an accuracy figure for our own tool either; the accuracy page explains why.