What it is
A free AI detector aimed at students, from a company whose main business is thesis editing and proofreading. Unlimited checks, no sign-up, 1,200 words per submission. A premium tier claims detection of GPT-4 output; the free one is described by Scribbr as handling GPT-2, GPT-3 and GPT-3.5 with “average accuracy”.
Scribbr is a brand of Learneo, Inc., which also owns QuillBot, Course Hero, CliffsNotes, LitCharts, Symbolab, Bartleby and LanguageTool. Hold on to the QuillBot part.
Nobody independent has ever tested it
This is the single most important fact on this page, and it is a negative, which is why you will not find it in any of the reviews currently ranking for this query.
Scribbr’s detector does not appear in Weber-Wulff et al. (2023), which tested fourteen tools. Nor in Perkins et al. (2024), which tested seven. Nor in Liang et al. (2023), which tested seven. Nor in the 2025 NBER working paper, which tested four. Every number you have ever read about how accurate Scribbr is came from Scribbr.
That does not make the tool bad. It makes the confidence around it unearned, and it means the honest answer to “how accurate is Scribbr’s detector” is: unknown, and unknowable from public evidence.
The comparison it runs, and wins
Scribbr publishes a test of twelve detectors that has become one of the most-cited sources in this space. Here are its results as it lists them:
| Tool | Accuracy |
|---|---|
| Scribbr (premium)— a Learneo brand | 84% |
| QuillBot— a Learneo brand | 78% |
| Scribbr (free)— a Learneo brand | 78% |
| Originality.AI | 76% |
| Sapling | 68% |
| Copyleaks | 66% |
| ZeroGPT | 64% |
| GPT-2 Output Detector | 58% |
| CrossPlag | 58% |
| GPTZero | 52% |
| Writer | 38% |
| OpenAI AI Text Classifier | 38% |
First, second and third are Scribbr premium, QuillBot and Scribbr free. Scribbr and QuillBot are the same company. The page says so nowhere.
It gets a little sharper. Two of the six categories the tools were tested on were “GPT-3.5 texts paraphrased by QuillBot” and “human-written texts paraphrased by QuillBot”. A third of the benchmark consists of text processed by the tester’s own sister product — the specific attack that its own detector places first against.
In fairness to Scribbr
We are not accusing anyone of faking a result, and there is a real case for the defence.
Scribbr describes its method openly — thirty texts, six categories of five, each 1,000 to 1,500 characters, human samples drawn from a thesis, an academic report, Wikipedia and online articles, AI samples generated from identical prompts. That is more disclosure than most vendors in this category offer, and it is the reason we can evaluate it at all. QuillBot is also a defensible choice of paraphraser precisely because it is what students actually use.
Its stated caveats are better than the industry norm, too: “AI detectors shouldn’t be treated as absolute proof that a text is AI-generated, but they can provide an indication in combination with other evidence,” and “even with advanced models, AI detectors may sometimes produce false positives by misidentifying human-authored text.” It even warns that “the companies behind some tools make strong claims about their reliability, but those claims are not supported by our testing” — a sentence that reads differently once you notice which companies were not being warned about.
The problem is not dishonesty. It is that thirty texts is a small sample in which one text moves a score by 3.3 points, the tester competes in its own benchmark, and the relationship between the top three entries is not disclosed. Those three facts together mean the table above should not be cited as independent evidence — by us, by Scribbr, or by the dozens of pages that currently do.
A smaller inconsistency
Scribbr’s detector page tells visitors that “the highest accuracy we found was 84% in a premium tool or 68% in the best free tool”. Its comparison page puts the best free tools — its own and QuillBot’s — at 78%. Most likely one page was updated and the other was not. We mention it only because it is a reminder that even a careful vendor’s figures drift across its own site, which is a decent argument for going to the primary page before quoting any of them.
What it is genuinely good for
A free, unlimited, no-signup self-check before you submit something is a real service, and Scribbr provides it without the dark patterns common in the category. If you want to see roughly how your own writing reads to this class of tool, it is a reasonable place to look. On the one measure that can be checked without trusting anybody’s accuracy claim — how much you are allowed to scan, how often, and whether you have to hand over an email first — it is the most generous of the free options we could find; the rest are set out in our ranking of the free detectors by what they actually let you check.
What it cannot do — what none of them can do — is establish that a piece of writing was or was not produced by a person. If a Scribbr score is being used against you, the fact that no independent study has ever evaluated this detector is a specific, checkable thing to put in front of whoever is deciding.
Where we stand
We sell a detector and a humanizer, which makes Scribbr a competitor and this page interested testimony. So, to hold ourselves to the standard we just applied: we have published no measured accuracy figure for our own detector either, and we will not until our benchmark runs under a method published in advance, against a corpus we did not choose after seeing the results. If we ever publish a comparison table with our own product at the top of it, apply everything on this page to us.
Compared against
- Scribbr vs QuillBot — Scribbr against QuillBot on what independent testing found, who each is sold to, and what happens to the text you submit.
- Scribbr vs Turnitin — Scribbr against Turnitin on what independent testing found, who each is sold to, and what happens to the text you submit.