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.
The 2025 NBER working paper tested both and put them in the same “secondary tier” — struggling on short passages and on text run through humanizing tools. The interesting result is what they share.
Last reviewed 16 August 2026 · The HumanFlow team
These are the two best-known detectors that a person can buy or use without an institution, and they are pitched at different buyers — GPTZero at educators and individuals, Originality.ai at publishers and agencies. They meet in the middle whenever someone wants a serious detector without a procurement process.
They also happen to have been tested together, which is unusual here. That gives the comparison a foundation most pairs in this category lack, and the foundation says something neither vendor's marketing does.
Something worth noticing about how this comparison is usually written elsewhere: nearly every version of it is published by one of the two companies, or by an affiliate paid on conversions. We are neither, and we are also not neutral — we run a detector. The mitigation is that everything below is either an independent finding with a citation or a vendor's own words, and both are labelled as such.
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 | GPTZero | Originality.ai |
|---|---|---|
| Who it is sold to | Consumers. An individual instructor can sign up between classes, which means the person reading your score may have no institutional guidance on how to read it — and your institution may not know the tool is in use. | Publishers, agencies and SEO teams first; education second and recently. Its CEO told Gizmodo in June 2024 that the company advised against academic use; by 2026 it markets an Academic Model for Educators and a Moodle plugin. |
| How you get at it | Free tier of 10,000 words a month, up to 10,000 characters in a single scan. | No free trial. Pro is $14.95 a month, or $12.95 a month billed annually, for 2,000 credits at 100 words each. |
| What it needs | No published minimum. The 2025 NBER working paper found it unsuitable for very short text. | Not published as a word floor. The NBER paper found it struggles on short passages. |
| What independent testing found | Recorded the highest false-positive probability of the fourteen tools in Weber-Wulff et al. (2023), at 50% against 0% for the best in that set. The 2025 NBER working paper placed it in a “secondary tier”, unsuitable for very short text and susceptible to humanizing tools. | Ranked second of four in the 2025 NBER working paper, in a “secondary tier” that struggles on short passages and on text run through humanizing tools. It was not among the fourteen tools in Weber-Wulff et al. (2023). |
| What the vendor claims | No accuracy figure is legible on its pricing page, which renders in the browser and serves $0/mo to an automated read. | 99%+ accuracy and false-positive rates between 0.5% and 1.5% depending on the model — published with no named corpus, no test-set size and no date. |
| What the vendor says about its limits | No published statement we could find. | Its terms forbid using the output as sole grounds for discipline. |
| What happens to your text | Not established from its published pages by us. | Its privacy policy states that unless you opt out, it may use your scan history and results to help train and improve its models. No fixed retention period is published. Institutional agreements exclude student data from general-purpose training. |
Every line above is summarised from our own examination of each tool, where the studies and vendor documents behind it are quoted and linked: GPTZero and Originality.ai.
Take the shared verdict first, because it outranks the difference. The NBER working paper placed both in a secondary tier below Pangram, and characterised that tier by two specific weaknesses: short passages, and text run through humanizing tools. If your use case involves either — a paragraph rather than an essay, or content someone has tried to disguise — the choice between these two matters less than the fact that both were found wanting there.
Where they separate is on false positives measured elsewhere. GPTZero recorded the highest false-positive probability of the fourteen tools in Weber-Wulff et al. (2023), at 50% against 0% for the best in that set. Originality.ai was not in that study, and publishes 0.5% to 1.5% for itself — a figure with no named corpus, no test-set size and no date, which makes it uncheckable rather than wrong.
The commercial models are opposites. GPTZero gives 10,000 words a month free, up to 10,000 characters a scan, and an individual instructor can be running it in a minute. Originality.ai has no free trial at all: Pro is $14.95 a month, or $12.95 billed annually, for 2,000 credits at 100 words each. So one is adopted casually and the other deliberately, which affects how much thought tends to go into how the output is used.
Their stated positions on academic use differ in a way worth weighing. Originality.ai's terms forbid using its output as sole grounds for discipline, and its CEO told Gizmodo in June 2024 that the company advised against academic use entirely, on false-positive grounds — before the company began selling an Academic Model for Educators. We found no equivalent published restriction from GPTZero, which is notable given that education is where it is most used.
The free-versus-paid split has an effect on the evidence base that is easy to overlook. GPTZero has been tested more than almost any detector partly because researchers could use it without a budget line, while Originality.ai's paywall makes it a harder tool to include in an academic study. Some of the difference in how much is known about these two is a difference in how easy each was to test, not in how much scrutiny each deserved.
The NBER paper is not peer-reviewed and says so, and one working paper is thin ground for any strong claim. It is still the only source that measured these two under the same conditions, which is why it does the work on this page.
Nothing here supports a ranking between them on false positives, because the one measurement of GPTZero's comes from a study Originality.ai was not in. Combining figures across studies with different corpora, different years and different tool versions produces a number that looks like a comparison and is not one.
Originality.ai's data handling belongs in the decision even though it is not an accuracy question. Its privacy policy states that unless you opt out, it may use your scan history and results to help train and improve its models, with no fixed retention period published. Institutional agreements exclude student data from general-purpose training. For GPTZero we could establish nothing either way from its published pages.
And one marketing claim needs correcting on Originality.ai's side. Google has said directly that it does not penalise content for being AI-generated — a spokesperson told Gizmodo it is “inaccurate to say Google penalizes websites simply because they may use some AI-generated content.” Its position is that low-value content produced at scale to manipulate rankings is spam however it was produced.
The shared tier finding is more useful than a ranking would be, and it is worth saying why. A ranking between two tools tells you which to buy. A shared weakness on short passages and humanized text tells you when to distrust either — which is the more actionable fact, because most disputed cases involve exactly the conditions the paper identified as the weak ones.
These two were built for different people and met in the middle by accident. GPTZero came out of the education panic of January 2023, built by a student, aimed at teachers, free at the door. Originality.ai came out of content marketing, aimed at publishers and agencies checking freelance copy, paid from the first scan with no free trial at all.
The result is two products with very different ideas of who the customer is, and the pricing shows it plainly. GPTZero gives 10,000 words a month free because its growth depends on being tried; Originality.ai charges $14.95 a month before you have scanned anything, because a publisher checking copy at volume is a different buyer from a teacher checking one essay.
Where they converge is in what the 2025 NBER working paper found, and the convergence is the interesting part. Two tools built for different markets by different kinds of company, tested on one corpus, landed in the same tier — struggling on short passages and on text put through humanizing tools. That suggests the limitation is a property of the approach rather than of either company's execution.
The most useful fact for either is the one they share: both were placed in a secondary tier by the only paper that tested them together, characterised by weakness on short passages and on text run through humanizing tools. If what was scanned was short, that finding applies directly and is worth citing.
If the result came from GPTZero, add the peer-reviewed record specific to it. Weber-Wulff et al. (2023) measured its false-positive probability at 50%, the highest of the fourteen tools examined. Used carefully that is not a claim of innocence but a claim about evidential weight — this tool's output cannot carry a finding on its own, and the largest study of these systems is the authority for saying so.
If the result came from Originality.ai, its own terms forbid using the output as sole grounds for discipline, and its CEO argued in June 2024 against academic use entirely because students submit too few essays for the false-positive risk to be acceptable. Quoting a vendor against its own product is more effective than quoting a competitor, which is what we are.
If two detectors disagreed about the same text, that is normal rather than evidence of malfunction, and it is worth saying so plainly. They are different classifiers with different training data and different thresholds. Disagreement tells you the measurement is less determinate than a percentage makes it look — which is itself relevant to whether a percentage should decide anything.
There is no basis for ranking these two on false positives. GPTZero's figure comes from a study Originality.ai was not in, and Originality.ai publishes its own figure with no corpus, no sample size and no date. Putting the two numbers side by side would produce a comparison that looks rigorous and is not.
The NBER working paper is not peer-reviewed and says so. It is the only source that measured both under the same conditions, which is why this page leans on it, and one working paper is a thin foundation for anything stronger than the tier-level claim it makes.
We could establish nothing about GPTZero's retention or training use from its published pages. Originality.ai documents that it may use scan history to train its models unless you opt out, and publishes no fixed retention period.
GPTZero's pricing cannot be captured — its page renders in the browser and serves $0/mo to an automated read, which is the worked example docs/10-competitor-capture.md was written around.
Pick GPTZero if you want a free check and will read the result as a rough signal. Its documented false-positive problem is a reason to discount a flag rather than to act on one.
Pick Originality.ai if you are checking commissioned or published content at volume — the publisher and agency workflow it was built for — and you have read its scan-history training clause. Not for disciplinary use: its own terms forbid that.
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.