Scribbr AI Detector: Accuracy, Word Limits, and the Open Questions

What Scribbr’s AI Detector Is

Scribbr’s AI Detector is a web tool that reads a block of text and estimates how much of it was machine-written. Scribbr the company is an Amsterdam-based academic-writing brand, best known for a citation generator, proofreading service, and plagiarism check; the AI detector is one member of that product family, and this page is about that one member only. When you paste text in, the tool returns a four-way read rather than a single AI-or-human verdict: it labels passages as AI-generated, AI-refined, human-refined, or human-written. That AI-refined category is Scribbr’s way of naming text that started as a model draft and was then edited by a person, which is the case a plain “AI likelihood” score tends to blur.

Scribbr’s Product Family (and the Plagiarism Checker It Isn’t)

The single most common mix-up worth clearing first: Scribbr’s plagiarism checker and its AI detector answer two unrelated questions, and searchers frequently arrive looking for the wrong one. A plagiarism check asks whether your wording matches an existing source somewhere. An AI check asks whether your wording reads as though a model produced it. Historically Scribbr’s plagiarism side has run in partnership with Turnitin, which matters later in this file, but the plagiarism product is not the subject here. There is one further gap in the family that drives real search traffic: people look for a “Scribbr AI humanizer” as if Scribbr sells one. It does not offer a dedicated humanizer; it offers a paraphrasing tool of a different kind. If rewriting a draft into your own register is what you actually came for, that is a separate job from checking one, and it lives elsewhere (our own tool does that side), but it is not something Scribbr’s detector page provides.

The Official Site, and an Impersonation Warning

Because the brand name trips people up in search (it is sometimes typed “scribber,” “scrbblr,” or “scrubber”), it is worth confirming plainly where the real tool lives before anything else. The official detector is at scribbr.com. A separate, unaffiliated site called scribbraichecker.com, which advertises an unverifiable “99.1% accuracy” figure of its own, is not Scribbr and is not covered by anything on this page; I name it once here only so you can tell the two apart and disregard the copycat.

Free vs Premium: Word Limits and Real Costs

The free tier is the fact most searches are chasing, so here it is without hedging: Scribbr’s free AI Detector accepts up to 1,200 words per submission, lets you run an unlimited number of those checks, and asks for no account. The exact sentence stating that limit sits verbatim in Scribbr’s own archived pages from October 18, 2024 through June 30, 2026, twenty-plus months unchanged (archived tool page, June 30, 2026; EV-scribbr-01). I quote the limit as Scribbr’s own stated term, which the vendor is the authoritative source for; I do not adopt the promotional tail it comes wrapped in.

The 500-to-1,200-Word History

Here is a piece of the record nobody else seems to have published dated and sourced. The free cap was not always 1,200 words. An archived snapshot of the same tool from July 3, 2024 shows the page’s own input-validation script enforcing a hard ceiling of 500 words; the script literally reads “Maximum of 500 words” (archived tool page, July 3, 2024; EV-scribbr-02). So the free limit more than doubled, from 500 to 1,200 words, sometime between July 3 and October 18 of 2024, and has held steady since. That is not a marketing claim on Scribbr’s part; it is a mechanical change readable directly in two dated copies of the page.

The 10,000-Word Thesis Problem

Now the honest limit that the 1,200-word number implies, because a real search behind this page is people carrying long documents. A 3,000-word paper will not fit one free submission, so you break it into three passes and read three separate scores that never combine into one. A 10,000-word thesis chapter turns into roughly nine passes, each judged on its own, and there is no whole-document reading waiting for you at the end; you assemble the picture yourself, in your head. Pasting a long paper in by hand, chunk after chunk, gets tiresome fast, so what usually happens is that a reader scans the opening pages, spot-checks one more block, and calls the job finished, leaving the untested middle a blind spot. I will be straight about my own side of this, because it would be dishonest to hide it: MeteGPT’s free anonymous tier tops out at 125 words a run, so it cannot swallow a 10,000-word thesis in one pass either. Neither free tool can. Those are two honest ceilings standing next to each other, and a long paper is a slog of repeated checks on both; any page that pretends its free tier makes the problem vanish is quietly selling you an upgrade.

What the Paid Version Actually Costs

There is no standalone premium AI detector to buy from Scribbr, a point most write-ups get wrong by implying a subscription exists. Paid AI detection is bundled into Scribbr’s Premium Plagiarism Check, and that check is priced per document by length rather than as a recurring plan: $19.95 for up to 7,499 words, $29.95 for 7,500 to 49,999 words, and $39.95 for 50,000 words and up, per check rather than by subscription (EV-scribbr-12; the bundling itself is Scribbr’s own stated term, EV-scribbr-03). So the model is pay-per-document, and the AI detector rides along with the plagiarism check rather than being sold on its own. The table below sets the two tiers side by side.

Table: Scribbr’s free and premium AI-detector tiers, showing word limits, per-check pricing bands, and signup. Prices and limits are Scribbr’s own stated terms as archived; vendor pricing can change without notice.

How Accurate Is Scribbr’s AI Detector?

“Is Scribbr’s AI detector accurate?” is the question this page exists to answer honestly, and the honest answer is layered: Scribbr publishes its own credited test, independent testers report mixed single-run results, and every figure below is attributed to whoever measured it rather than laundered into a flat number. I am not going to substitute my voice for any of these sources’ disclosed results.

Scribbr’s Own Test, Credited

Scribbr is one of the few vendors in this space that publishes a disclosed method instead of a bare headline, and that is worth crediting even while flagging the obvious self-interest. Scribbr’s own comparison, authored by co-founder Koen Driessen (named and credentialed), describes its design directly: 30 texts across six categories (human-written, GPT-3.5, GPT-4, mixed human-plus-AI, and paraphrased AI, five texts each, 1,000 to 1,500 characters apiece), run across 12 detectors. Scribbr’s own testing reports its Premium AI Detector at 84% overall accuracy, the best paid result in its set, and its free detector at 78%, tied for best free (archived comparison, August 11, 2025; EV-scribbr-08). Read those as Scribbr’s own numbers on Scribbr’s own product, not as an independent ranking. Scribbr also discloses a caveat against its own interest, which is the part I most respect: it reports that its detection of paraphrased text runs only around 60%, well below its headline figure, a vendor naming the exact case where its tool is weakest.

There is a companion point from Scribbr’s own explainer, credited to Jack Caulfield and revised through July 2025. It describes detectors as reading two signals, perplexity and burstiness (how predictable the wording is and how much that predictability varies), and states plainly that no AI detector can reach complete accuracy (archived explainer, revised July 31, 2025; EV-scribbr-09). That admission, from the vendor itself, is the correct frame for every number on this page.

What Independent Testers Found

Two dated third-party tests are worth reporting precisely, each with its conflict of interest attached, and neither one generalisable into a rate. AI Busted, a site that sells a competing detector and humanizer, ran 20 human-written college essays (real essays submitted before ChatGPT existed, with no AI involvement) through Scribbr’s detector, and its own test found 4 of those 20 flagged as partly AI-written (AI Busted’s Scribbr review, May 21, 2026; EV-scribbr-10). Read that as exactly what it is: 4 flags out of the 20 essays in that one test, from a source with a product to sell, on a specific dated run. It is not a false-positive rate that generalises to your essay, and it cannot be aggregated with anything else here; the sample and conditions do not travel.

Pulling the other direction, Originality.ai, itself a rival detector, ran four ChatGPT-generated marketing samples through Scribbr and reported Scribbr scoring 0% AI on three of the four (a blog post, product description, and promotional email) and 95% on the fourth, a cover letter, while its own tool flagged all four (Originality.ai’s Scribbr review, October 28, 2025; EV-scribbr-11). That is a single dated data point on false negatives for AI marketing copy specifically, from a source with a direct interest in outscoring the tool it is reviewing, on a sample of four. Treat it as that and nothing larger.

Where Their Own Numbers Disagree

For anyone auditing the tool rather than just using it, Scribbr’s public record contradicts itself in a few places worth logging without editorialising. Its AI-detector FAQ lists four supported languages (English, Spanish, German, French); its plagiarism FAQ lists five, adding Dutch. Same brand, two different counts. An older 2023 methodology page cites roughly 60% as an average detector accuracy, while the 2024 comparison spreads its detectors from 52% to 84%, and the two figures are never reconciled on the site. The live 2026 FAQ still repeats figures from a November 2024 revision, so a reader landing today is reading numbers dated to an earlier version of the tool. None of that is an accusation; it is a set of dated inconsistencies in one vendor’s own copy, recorded here because a careful reader should know the numbers on the site do not fully agree with each other.

Who Built Scribbr’s AI Detector? An Open Question

This is the section with the most genuinely unresolved question, and the discipline it needs is restraint: I will use only the verbs says, said, claims, removed, and disputed, because I do not know the answer and neither, on the public record, does anyone else. There are three dated positions on what powers Scribbr’s AI detector, and I am going to lay all three out and then decline to pick one.

Three Dated Claims, No Verdict

The first position is Scribbr’s current one. Scribbr’s Premium Plagiarism Check page, in a snapshot from June 3, 2026, says the bundled AI detector runs on Scribbr’s own proprietary software, describing it as an in-house classifier (archived plagiarism-checker page, June 3, 2026; EV-scribbr-03). Take that as Scribbr’s own present claim about itself.

The second position is Scribbr’s own earlier one, and it does not square cleanly with the first. An archived version of Scribbr’s AI-detector page from February 1, 2024 said the Premium AI Detector was “powered by Turnitin,” and separately listed Scribbr as an “Authorized partner of Turnitin” (archived tool page, February 1, 2024; EV-scribbr-04). That the page said this on that date is archive-verifiable; whether Turnitin ever actually powered the detector remains Scribbr’s own claim from the time, which is why I phrase it as what the page stated rather than as a fact about the engine.

The third position is a neutral, dated change. The next archived snapshot of that same page, from March 7, 2024 (34 days later), no longer describes the Premium AI Detector as powered by Turnitin; the Turnitin partnership was rewritten to scope only to plagiarism matching, and the AI-detection-specific Turnitin wording does not reappear in any later snapshot checked through June 30, 2026 (archived tool page, March 7, 2024; EV-scribbr-05). The wording changed between two dated snapshots; that is assertable as fact. Why it changed is not something these snapshots reveal, and I attach no motive to it, since a partnership can be renegotiated or rescoped for reasons that have nothing to do with what powers the tool.

A Disputed Fourth Claim, and the One Fact Underneath It

There is a fourth claim in circulation, and it needs its disputed label carried the whole way. A dated review by GPTZero, itself a competing detector, which recommends itself later in the same article without disclosing the conflict, claims Scribbr’s detector is a “wrap-around” for QuillBot’s engine, on the grounds that both products share a parent company (GPTZero’s Scribbr review, January 7, 2025; EV-scribbr-06). I register that as a disputed, competitor-sourced claim, not as fact, because the specific technical assertion, that the two run the same engine, is not corroborated by any neutral source I could verify this run.

What is independently corroborated is the ownership fact underneath it, and only that. Learneo’s own December 2022 press release (Learneo is the parent company, formerly Course Hero) names Scribbr and QuillBot among its operating businesses, alongside CliffsNotes, LitCharts, and Symbolab (Learneo press release, December 14, 2022; EV-scribbr-07). So common ownership is a settled fact; a shared detection engine is not. The distance between those two is exactly where the disputed claim lives, and I leave it there.

One more factual note for the instructor or reviewer auditing scope: Scribbr’s live 2026 hero copy names ChatGPT, Copilot, and Gemini as the models its detector targets. It does not name Claude or any GPT-5-class model. I state that as a scope fact, not a gotcha; a detector’s named coverage is simply worth knowing before you trust its read on newer output.

False Positives and ESL Writers

A false positive (genuine human writing wrongly labelled as machine-written) is the failure that actually harms a student, and it falls hardest on one group. I want to be precise about the evidence here, including about what it does and does not measure.

What ESL Writers Should Expect

The strongest evidence on this risk is a peer-reviewed Stanford study, and its finding is stark: commercial GPT detectors, tested on real TOEFL essays written by non-native English speakers, wrongly labelled more than half of that genuine human work as AI-generated, an average false-positive rate of 61.3%, while flagging native-speaker essays far less often (Liang et al., Patterns, Cell Press, DOI 10.1016/j.patter.2023.100779; EV-best-ai-humanizer-01). One point of honesty that matters: that study tested detectors as a class, and Scribbr’s own detector was not among the specific tools it examined. So I am not presenting 61.3% as a Scribbr measurement, because it is not one. What it is, is a documented empirical floor for the whole class of perplexity-based detectors, and Scribbr’s own explainer (above) confirms Scribbr’s tool reads exactly that kind of signal. The class-level warning applies; the specific number is the class’s, not Scribbr’s.

The mechanism is the reason the risk concentrates the way it does. Non-native writers often lean on more regular sentence patterns and a more common vocabulary, and to a detector that scores predictability, “regular and clear” can read as “machine-made.” That is not the detector catching AI; it is the detector catching a writing style that happens to correlate with second-language English, and mislabelling it. If you are an ESL writer flagged on your own work, that flag is consistent with a measured, published bias rather than proof you did anything wrong. The practical protection is the same on every tool: keep your drafts, your outline, and your notes, because your writing process is the record a style score cannot argue with. If you want to check where a draft stands against a differently-built engine before anyone else does, our own detector runs free with no account, and it is candid about the identical blind spot on its own page.

Scribbr vs QuillBot: One Note

Because the two brands share a parent, one practical question comes up often enough to answer directly and once. Scribbr and QuillBot are commonly owned under Learneo (EV-scribbr-07), and a competitor claims they share a detection engine, a claim that remains disputed and uncorroborated, as covered above (EV-scribbr-06). The practical takeaway does not depend on settling that dispute: if you run your text through Scribbr and then through QuillBot hoping for a second, independent opinion, you should treat that pairing with caution, because the common ownership means you cannot assume the two are genuinely separate reads the way two unrelated detectors would be. For a second opinion you can be confident is independent, reach for a detector built by an unrelated company on different technology.

Is Scribbr’s AI Detector Worth It? The Verdict

The plain read, after all of that: Scribbr’s free AI Detector is a reasonable, no-cost first-pass check with a clearly-documented 1,200-word limit and, to Scribbr’s credit, a published test method that most rivals do not bother to disclose. Its own credited research reports respectable accuracy on clean AI text and openly concedes a weak spot on paraphrased writing. What no honest reading supports is treating it as a settled authority: its provenance is unresolved on its own dated record, its own numbers disagree with each other in places, and the false-positive risk that runs through this whole category, heaviest for second-language writers, is not one Scribbr’s class of tool has escaped. Use it as one dated data point, read its result as a signal rather than a ruling, and if you are an ESL writer, weight it accordingly.

Common Questions

Is Scribbr’s AI detector accurate? By its own credited testing, Scribbr reports 84% for its Premium detector and 78% for the free one across a disclosed 30-text set (EV-scribbr-08), while conceding its own detection of paraphrased text runs only around 60%. Read those as the vendor’s own numbers, not an independent ranking, and remember that no detector, by Scribbr’s own explainer, reaches complete accuracy (EV-scribbr-09).

How does Scribbr’s AI Detector work? Scribbr’s own explainer describes it as reading perplexity and burstiness (how predictable your wording is and how much that predictability varies) and returning a four-way read: AI-generated, AI-refined, human-refined, or human-written (EV-scribbr-09). It scores the statistical shape of writing, not its meaning.

Is Scribbr’s AI checker free? Yes. The free tier reads up to 1,200 words per submission with unlimited checks and no signup (EV-scribbr-01). Paid AI detection is not sold separately; it is bundled into the Premium Plagiarism Check at $19.95 to $39.95 per document by length (EV-scribbr-12).

Is Scribbr’s AI detector the same as QuillBot’s? They share a parent company, Learneo (EV-scribbr-07). A competing detector claims they also share an engine (EV-scribbr-06), but that specific claim is disputed and uncorroborated by any neutral source, so it is safest to assume the two are commonly owned and to seek a genuinely independent second opinion from an unrelated tool.