HumanFlow is not one of the two tools on this page. We sell a rewriter and run a detector of our own, so weigh this accordingly — our methodology sets out how we source what appears below and what we refuse to claim.
QuillBot states that when a result is unclear its model leans towards calling text human. Copyleaks sells a sensitivity setting that does the opposite. They are tuned in opposite directions, on purpose.
Last reviewed 16 August 2026 · The HumanFlow team
Both offer a detector, one free and one paid, and both are reached by people who want a second opinion on a document. What makes the pair worth a page is that they have made opposite decisions about which kind of error to make — and both have said so publicly, which is rare enough to be useful.
There is also a real asymmetry of purpose behind them. Copyleaks sells detection to institutions; QuillBot's detector is a free surface on a paraphrasing business. That shapes who each is trying not to annoy.
The conflict of interest on this page runs deeper than usual and should be visible before the argument. QuillBot's main business is paraphrasing, which is the market we sell into; Copyleaks sells detection, which is the other half of what we do. We are a competitor to both, in different ways. Every figure below is theirs, linked, and the reasoning is set out so it can be disagreed with.
No prices. Neither of these is a tool most readers choose on cost, and for one of them there is no public price at all — detector pricing, where it could be captured, is on our detector pricing page.
| Dimension | Copyleaks | QuillBot |
|---|---|---|
| Who it is sold to | Institutions and individuals both — one of the few detectors genuinely sold to universities rather than only to people. Education is priced on full-time student count, with integrations for Canvas, Moodle, D2L Brightspace, Blackboard, Schoology, Sakai and Edsby. | Consumers. The detector is free; the paraphrasing suite is the product. Owned by Learneo, Inc., the same parent as Scribbr. |
| How you get at it | Paid. Personal is $16.99 a month, or $13.99 a month billed annually; one credit covers up to 250 words. | Free on its site, with no published scan cap. |
| What it needs | Not published as a word floor. Its highest sensitivity setting is described as designed to flag text put through a humanizer or spinner. | A minimum of 80 words, with its own page advising that 300 or more produces more reliable scores. |
| What independent testing found | Best of the seven tools in Perkins et al. (2024) — and it still missed 39% of the AI cases and produced the highest false-accusation rate in that set. Its accuracy fell from 73.9% to 58.7% once simple adversarial techniques were applied. | None specific to it. It cites RAID, an independent benchmark, for a detection rate — but RAID measures how much AI text a tool catches, not how often it flags human writing. |
| What the vendor claims | 99.97% accuracy and a 0.026% false-positive rate on its V10 model. | A 99% detection rate, citing RAID. That is a detection rate, not a false-positive rate; they publish the first and not the second. |
| What the vendor says about its limits | No published statement we could find. | States that when the result is unclear the model tends to classify text as human-written, names heavily paraphrased writing and formulaic legal and academic prose as what it gets wrong, and says never to rely on an AI detector alone. |
| What happens to your text | Paid plans include the Shared Data Hub, which Copyleaks describes as a library of user-submitted documents that scans are compared against. Its privacy policy states it uses this information to train its models, with opt-out available to direct customers by contacting support. | Not established from its published pages by us. |
Every line above is summarised from our own examination of each tool, where the studies and vendor documents behind it are quoted and linked: Copyleaks and QuillBot.
QuillBot's threshold choice is the most informative disclosure either company makes. Its documentation states that when the result is unclear, the model tends to classify text as human-written, which reduces false positives. The practical consequence is precise: a “human” verdict from QuillBot carries less information than an “AI” one, because the tool is built to resolve doubt in that direction. Very few vendors state a threshold choice at all.
Copyleaks is tuned the other way at its top setting, and says so. Its highest sensitivity is explicitly described as designed to flag AI text run through a humanizer or spinner — which is a setting that accepts more false positives in exchange for catching more disguised text. An institution turning that on is making a policy choice about which error it prefers, whether or not anyone frames it that way.
QuillBot is also unusually specific about what it gets wrong, and the answer should worry any academic writer. Its own page names heavily paraphrased writing and highly formulaic content such as legal jargon and academic writing. That is a vendor stating that careful, conventional academic prose is more likely to be misread by its tool — the same population every independent study of this category has flagged.
Commercially they barely overlap. Copyleaks is paid, sold on full-time student count to institutions with integrations for Canvas, Moodle, D2L Brightspace, Blackboard, Schoology, Sakai and Edsby, with Personal at $16.99 a month or $13.99 billed annually. QuillBot's detector is free with no published scan cap and a floor of 80 words, with its own page advising 300 or more for a reliable score; the paid product there is the paraphrasing suite.
The two threshold choices produce mirror-image failure modes, and knowing which you are exposed to is more useful than knowing an accuracy rate. Copyleaks at high sensitivity will flag human writing that happens to be conventional and predictable — the false positive. QuillBot will pass AI text it is unsure about — the false negative. If you are the person being assessed, only one of those errors can hurt you, and it is not the same one for both tools.
Copyleaks has been tested independently and QuillBot's detector has not. In Perkins et al. (2024) Copyleaks was the best of seven tools — while missing 39% of the AI cases and producing the highest false-accusation rate in that set, with accuracy falling from 73.9% to 58.7% under simple adversarial techniques. Its own figures are 99.97% accuracy and a 0.026% false-positive rate on its V10 model.
QuillBot cites RAID, which is a real independent benchmark, for a 99% detection rate. That is a detection rate: it measures how much AI text the tool catches. It says nothing about how often the tool flags human writing, which is the number that matters to anyone at risk of being accused. They publish the first and not the second, and the distinction is not incidental — it is the difference between the two error types these tools trade off.
So the comparison comes down to a tool with a mixed independent record and a tool with a candid account of its own biases and no independent record at all. Both positions are defensible for different readers, and neither supports treating an output as proof. QuillBot says as much itself: its documentation states never to rely on an AI detector alone when reviewing results.
QuillBot's disclosure about formulaic academic prose lines up with the strongest finding in the independent literature, which is why it is worth more than a vendor caveat usually is. Liang and colleagues found detectors misclassify non-native English writers' work at dramatically higher rates, for the same underlying reason: careful, conventional, vocabulary-constrained prose reads as statistically ordinary. A vendor and a peer-reviewed study naming the same weakness is about as close to corroboration as this field offers.
Copyleaks sells detection to organisations that will act on the results — universities priced per full-time student, enterprises with compliance obligations. When a customer will make decisions about people using your output, sensitivity becomes a setting rather than a default, and a documented position on humanized text becomes something the sales process requires.
QuillBot's detector is free and its paraphrasing suite is the business. The detector brings people to a site where the product is a rewriting tool — which is a genuinely awkward pairing, and QuillBot's own documentation is unusually candid about the seam: it names heavily paraphrased writing as something its detector gets wrong, on a site that sells paraphrasing.
The tuning follows the business in each case, and they point in opposite directions. Copyleaks offers a top sensitivity setting explicitly aimed at humanized and spun text — accepting more false positives to catch more disguised writing. QuillBot resolves unclear cases towards human, reducing false positives at the cost of missing AI text. Both are defensible engineering choices and neither is neutral.
If the result came from Copyleaks, ask which sensitivity setting produced it. The highest one is designed to flag text run through a humanizer or spinner, which means it is deliberately tuned to accept more false positives — and an institution that enabled it made a policy choice about which error it would rather make. You are entitled to know that choice was made and by whom.
The independent record on Copyleaks is more useful than its marketing in this conversation, and it cuts both ways honestly. In Perkins et al. (2024) it was the best of seven tools tested — while missing 39% of the AI cases and producing the highest false-accusation rate in that set, with accuracy falling from 73.9% to 58.7% under simple adversarial conditions. Best available and safe to act on are different claims.
If the result came from QuillBot, the direction of its threshold matters to how you read it. Its own documentation states that when a result is unclear the model leans towards classifying text as human — so an AI verdict from QuillBot is a stronger signal than a human verdict from it, and being flagged by the permissive tool is worth taking seriously as feedback about how predictable the prose is.
QuillBot also names what it gets wrong, and it is the most useful sentence on either vendor's site for an academic writer: heavily paraphrased writing, and highly formulaic content such as legal jargon and academic writing. That is a vendor stating that careful conventional academic prose is what its tool most often misreads.
Only one side of this pair has been independently tested. Copyleaks appears in Perkins et al. (2024); QuillBot's detector appears in no independent study we hold, so there is no basis for ranking them against each other on accuracy.
QuillBot cites RAID for a 99% detection rate, and RAID is a real independent benchmark. But a detection rate measures how much AI text a tool catches and says nothing about how often it flags human writing — and we could find no false-positive rate for QuillBot from any source. That is the number that matters to anyone at risk of a wrong flag, and it is the one not published.
We could not establish whether QuillBot uses submitted text for training or how long it retains it. Copyleaks documents both, which is what makes its arrangement open to criticism at all.
Copyleaks's prices were read on 15 August 2026 and could not be re-verified on 17 August, when its pricing page began returning a Cloudflare challenge to automated requests. QuillBot publishes no detector price because the detector is free.
Pick Copyleaks if you are an institution and need something to actually deploy — integrations, a sensitivity policy, and independent testing behind it. Read what the Shared Data Hub does with submissions before turning it on.
Pick QuillBot if you want a free read and, more usefully, the clearest published account of how a detector's threshold is set. Knowing that unclear cases resolve towards “human”, and that formulaic academic prose is a named weak point, tells you how to read a score.
A detector reports how statistically machine-typical a piece of prose reads. It has no access to how the text was produced, so it cannot establish authorship in either direction — a flag is not evidence of AI use, and a clean result is not a clearance. If you have been accused on the strength of one, drafting history is what actually answers it. And if your institution requires you to disclose AI assistance, disclose it — nothing on this page changes that obligation.