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.
These do not do the same job. Turnitin guesses at the product; Authorship records the process. Only one of them is evidence about what actually happened.
Last reviewed 16 August 2026 · The HumanFlow team
Students meet this pair from opposite ends. Turnitin is what runs on the finished file after submission, chosen by the institution. Authorship is something you switch on before you start, in your own editor, and it exists precisely because a detection score is weak evidence. Comparing them is really comparing two answers to the question “how would you show you wrote this”.
It is also the pair that comes up in appeals. When a Turnitin score is the accusation, the useful question is what could answer it — and a provenance log is the strongest thing available short of full version history.
Before the comparison, the honest disclaimer about scope: these two are not substitutes and nobody chooses between them. You will meet Turnitin because your institution runs it, and you will have Authorship only if you turned it on. So this page is less a recommendation than an explanation of what each kind of evidence can support — which is the question that actually matters when one of them is being used to decide something about you.
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 | Turnitin | Grammarly Authorship |
|---|---|---|
| Who it is sold to | Institutions only. A student or an individual instructor cannot buy it; it arrives inside the systems a university already licenses. | Individuals and organisations, as part of Grammarly Pro. It is a feature of a writing tool rather than a product an institution points at submitted work. |
| How you get at it | No public checker. If your institution has not enabled the indicator for your class, there is no way to see your own score. | Turned on by the writer, in their own editor. Nobody can run it on your document after the fact — which is the whole difference between it and everything else on this site. |
| What it needs | At least 300 words of prose, up to 30,000. .docx, .pdf, .txt or .rtf under 100MB, in English, Spanish or Japanese. | Not applicable. It records provenance as a document is built rather than judging finished prose. |
| What independent testing found | Scored highest of the fourteen tools in Weber-Wulff et al. (2023) — in a study whose own conclusion was that detection tools “are neither accurate nor reliable”, with every tool below 80% accuracy. | None on Authorship, because it is not a classifier and there is nothing to benchmark. Grammarly's separate AI detector cites RAID. |
| What the vendor claims | Aims to keep false positives under 1% above a 20% detected share. Below 20% it attributes no score at all, showing an asterisk, because its own testing found more false positives in that band. | For its separate AI detector: 99% detection accuracy and first place on the RAID benchmark, both quoted from its own page. |
| What the vendor says about its limits | States its indicator is not intended as the sole basis for an academic misconduct finding. | Unusually, the same site states repeatedly that no AI detector is 100% accurate and that no detector can conclusively determine whether AI was used. |
| What happens to your text | Submissions may be retained in its repository depending on the institution's configuration, which the student does not set. | It creates a provenance record of your own drafting. That record is evidence you can produce — and evidence that can be demanded of you. |
Every line above is summarised from our own examination of each tool, where the studies and vendor documents behind it are quoted and linked: Turnitin and Grammarly Authorship.
Turnitin reads finished prose and estimates what share of it is statistically machine-typical. It never sees the writing happen. It needs 300 to 30,000 words in .docx, .pdf, .txt or .rtf under 100MB, in English, Spanish or Japanese, and it attributes no score at all between 0% and 20% because its own testing found more false positives in that band.
Grammarly Authorship categorises text by where it came from as you write — typed by you, pasted from an online source, or generated by AI — so you can show how a document was produced. It records provenance rather than inferring it afterwards. Grammarly describes it as part of Grammarly Pro.
That difference decides who controls the evidence, which is the practical heart of this comparison. You cannot run Turnitin on your own work unless your institution enables it, and you cannot see the report unless someone shows it to you. Authorship is yours: you turn it on, and the record is something you can produce. It is also something that can be asked for, and both of those are true at once.
Neither is proof, and Authorship's limits are specific rather than general. It records what happened in one editor. Text typed elsewhere and pasted in appears as pasted, which is indistinguishable from pasting from a model. Work done outside Grammarly is not covered at all. Treat it as one strong item in a drafting record, not as a certificate.
There is an equity dimension to provenance logging that deserves stating, because it decides who benefits. Authorship protects the writer who turned it on before they needed it — which favours students who knew the risk existed, had a Grammarly Pro subscription, and drafted in one editor. A student writing in a shared computer lab, across a phone and a library machine, has no comparable record available and has done nothing wrong.
Turnitin scored highest of the fourteen tools in Weber-Wulff et al. (2023), in a study concluding that detection tools “are neither accurate nor reliable” with every tool below 80% accuracy. Being the best of that set is a real result and a limited one.
There is no equivalent number for Authorship because there is nothing to benchmark. It is not a classifier and makes no prediction, so it has no accuracy rate, no false-positive rate and no study. That is a category difference, not a gap in the evidence.
Grammarly's separate AI detector — a different feature from Authorship — claims 99% detection accuracy and first place on the RAID benchmark, both from its own page. Alongside those, and unusually for this category, the same site states repeatedly that no AI detector is 100% accurate and that no detector can conclusively determine whether AI was used.
The general point survives both tools. A detector cannot establish authorship in either direction, and the only technologies that record what actually happened are provenance logging and watermarking — one of which covers a single editor and the other a narrow slice of text.
Neither approach escapes the underlying problem, and it is worth naming what that problem is. A detector infers process from product and gets it wrong at a measurable rate. A provenance log records process, but only inside one application, so it establishes what happened there and stays silent about everything else. Both leave a gap, and in both cases the gap is filled by a human being deciding what to believe.
Turnitin is an integrity company selling to institutions, and its AI indicator is the newest layer on a product built to answer a question institutions were already asking. Grammarly is a writing company selling to individuals, and Authorship is the newest layer on a product built to help people write. Neither company set out to answer the question this page is about; both arrived at it from businesses that already existed.
That difference in starting point produced two genuinely different technologies rather than two versions of one. If your business is judging finished documents, you build a classifier that reads finished documents. If your business lives inside the editor while someone types, you can record what happened instead of inferring it afterwards — and recording is a strictly better kind of evidence when it is available.
It also explains who controls the output, which is the part that matters most in practice. Turnitin's customer is the institution, so the institution sees the score and the student may not. Grammarly's customer is the writer, so the writer holds the record and decides whether to produce it. Same subject matter, opposite arrangements of power over the evidence.
If a Turnitin score is the accusation, the useful response is evidence about process rather than argument about the number. Version history in Google Docs or Word, notes, outlines, earlier drafts, the sources you opened — these speak to how the document was produced, which is exactly what a detector cannot observe at any accuracy level.
Two facts from Turnitin itself are worth having. It states its indicator is not intended as the sole basis for an academic misconduct finding, and it shows an asterisk rather than a figure below 20% because its own testing found more false positives in that band. If you have been handed an asterisk and told it is a low score, that is a misreading of the tool's own design.
Authorship helps only if it was running while you wrote, which is the whole difficulty with it. It cannot be produced after the fact, and that is the argument for turning it on before you need it rather than a reason to dismiss it. If it was running, the log covers what was typed, what was pasted and what was generated, in one editor.
Be clear-eyed about its limits when you produce it. Text composed elsewhere and pasted in appears as pasted, which is indistinguishable from pasting from a model, and work done outside Grammarly is not covered at all. It is a strong item in a drafting record rather than a certificate, and presenting it as more than that invites the objection.
There is no accuracy comparison to make here, and the reason is a category difference rather than a gap in the literature. Authorship is not a classifier and issues no prediction, so it has no accuracy rate, no false-positive rate and nothing to benchmark. Comparing it with Turnitin on accuracy is a question with no coherent form.
Grammarly's separate AI detector — a different feature from Authorship — claims 99% detection accuracy and first place on RAID, both from its own page. We hold no independent verification of either, and no false-positive rate for it at all.
We could not establish how an institution treats an Authorship log in practice. Whether a disciplinary panel accepts it, weighs it, or has ever seen one is a question about institutional custom, and no data we hold answers it.
Nor could we establish what Grammarly retains from an Authorship record, or for how long. That matters in both directions, because the same log is evidence you can produce and evidence that can be demanded of you.
We could not establish whether any institution formally accepts a provenance log as evidence in a misconduct process. Grammarly documents the feature and Turnitin documents the indicator; neither documents how a panel should weigh one against the other, and that is the question a student in trouble actually needs answered.
Pick Turnitin if the choice is not yours. Turnitin is what an institution runs, and understanding the 20% suppression band and the 300-word floor is more useful than any preference between the two.
Pick Grammarly Authorship if you are the writer and you want a record. If your work may be questioned, provenance logging is the strongest evidence available short of full version history — provided you are comfortable having created something that can be demanded of you.
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.