Detector companies advertise 98–99% accuracy. Peer-reviewed studies, OpenAI and several universities tell a very different story. Here’s the evidence, claim by claim.
Evidence review: this verdict is built from published studies, official statements and vendors’ own pages, all linked below and checked on 14–15 Oct 2026. Hyptor Lab’s own hands-on test is scheduled and will be added to this page. No affiliate links. How we test
The short answer
Not reliably. Independent studies found detectors far less accurate than advertised, easy to fool, and prone to flagging non-native English writers. Treat a score as a hint, never as proof.
Use them for: a prompt to look closer at your own writing · Never use them for: accusing someone · Confidence: Medium — strong, consistent research, but the key studies are from 2023–24 and tools have changed since.
What detectors claim vs what the evidence shows
| Their claim | Verdict | What the evidence shows |
|---|---|---|
| “99% accuracy” (GPTZero, Grammarly), “over 99%” (Copyleaks), “99% detection rate” (QuillBot), “98.4%” (ZeroGPT) [3–7] | Overstated | A peer-reviewed test of 14 tools found “all scored below 80% of accuracy and only 5 over 70%” [1]. Another found 39.5% mean accuracy on unedited AI text [2]. Vendor figures come from their own benchmarks. |
| Very low false positives: “.03%” (Copyleaks), “<1%” (ZeroGPT) [4][7] | Overstated | Independent false-positive rates ranged from 0% to 50% across tools [1]; one study saw Copyleaks flag 5 of its 10 human samples [2] (a small sample). |
| Turnitin: “false positive rate of less than 1% for documents containing more than 20% AI-generated content” [9] | Holds, with caveats | Turnitin had 0% false positives in one independent test [1], but Turnitin itself reports “a higher incidence of false positives” below 20% AI and ~4% at sentence level [16]. |
| Detectors can catch AI text that’s been edited or paraphrased | Doesn’t hold | Accuracy on machine-paraphrased text was 26% [1]; simple evasion cut accuracy by 17.4 points on average [2]; RAID found detectors “easily fooled by adversarial attacks” [14]. |
Why detector accuracy is a fairness problem
A missed AI essay is an annoyance. A human essay flagged as AI can mean a failed course or a lost client. That’s why the number that matters most isn’t “accuracy” — it’s the false positive rate: how often real human writing gets labelled as AI.
Vanderbilt University did the maths when it switched off Turnitin’s detector in 2023: even at Turnitin’s own 1% rate, “around 750 student papers could have been incorrectly labeled” out of the 75,000 it received in 2022 [12].
How we built this verdict
- Sources: three peer-reviewed studies [1][2][10], one major benchmark [14], official statements from OpenAI, Turnitin and two universities [11–13][16], and each vendor’s own claims page [3–9].
- Rule: every number was checked on the source page itself; anything we couldn’t open was left out.
- Limits: the independent studies tested 2023–24 versions of these tools. We found no peer-reviewed 2025–26 comparison of all seven. Hyptor Lab’s own test (90 known texts × 6 detectors) will update this page.
Independent tests found accuracy well below the 99% claims
Finding: Weber-Wulff and colleagues tested 14 tools, including GPTZero, ZeroGPT and Turnitin, and concluded they are “neither accurate nor reliable” [1]. Perkins and colleagues ran 805 tests and found a mean accuracy of 39.5% on unedited AI text, with GPTZero the weakest at around 26% [2].
In practice: a “99%” badge tells you what a company measured on its own test set — not what will happen to your essay.
Non-native English writers get flagged far more often
Finding: Stanford researchers ran 91 TOEFL essays written by non-native speakers through 7 detectors. The average false-positive rate was 61.3%, and 97.8% of the essays were flagged by at least one detector. Essays by US eighth-graders were classified accurately [10].
In practice: simpler vocabulary looks “AI-like” to these tools. GPTZero and Turnitin have published their own research disputing this bias [2]; independent replication with current tools is still needed.
Light editing defeats most detectors
Finding: on machine-paraphrased AI text, accuracy fell to 26% [1]. Simple evasion techniques cut accuracy by an average of 17.4 points, with Turnitin dropping the most (42.1 points) [2]. The RAID benchmark — over 6 million generations — found “current detectors are easily fooled by adversarial attacks” [14].
In practice: detectors mostly catch people who paste raw AI output — and occasionally innocent people.
The surprise: even the makers don’t trust them
OpenAI withdrew its own AI classifier in July 2023 “due to its low rate of accuracy” — it caught 26% of AI text and wrongly flagged human text 9% of the time [11]. And the detector companies’ own fine print agrees with the critics: GPTZero says results “should not be used to punish or as the final verdict” [3]; QuillBot says “never rely on AI detection alone” [6]; Turnitin says it “does not make a determination of misconduct” [9].
Where AI detectors genuinely help
They’re reasonable at spotting long, unedited, pasted AI text — Turnitin and Copyleaks detected around 61–65% of it in one study [2], and Turnitin produced no false positives in another [1]. Used as a nudge to have a conversation, review drafts or check version history, a score can be one useful data point among several.
What they cost
| Detector | Free tier | Cheapest paid |
|---|---|---|
| GPTZero | 10,000 words/month | Premium (USD price not shown in our region) |
| ZeroGPT | 15,000-character box | $8.99/mo billed annually; $14.99 monthly |
| QuillBot | Up to 1,200 words per scan | $4.17/mo billed annually |
| Copyleaks | Up to 25,000 characters | $13.99/mo ($167.88/yr) |
| Scribbr | Up to 1,200 words per submission | — |
| Grammarly | Free detector page | Pro, listed at $12 |
| Turnitin | Institutions only | |
What to use instead — or alongside
Teachers: the University of Waterloo stopped using Turnitin’s AI detection in 2025, saying “there are no products on the market that can effectively identify GenAI” [13]. Drafts, version history, in-class writing and a conversation are more reliable evidence. Students: keep your drafts and version history — they’re your best defence against a false flag. Editors: judge the work, not the score.
Questions people ask
Are AI detectors accurate?
Not reliably. Independent studies found accuracy below 80% for every tool tested [1] and much lower on edited text [2]. Scores are hints, not proof.
What is the best AI detector for teachers?
No detector is accurate enough to base a decision on — vendors say so themselves. Turnitin had the fewest false positives in one independent test [1], but several universities have disabled it [12][13].
Is GPTZero the best AI detector?
Its 99% claim isn’t backed by independent tests: one study measured around 26% detection [2] and another a false-positive rate of up to 50% [1]. Those tests used 2023–24 versions.
Can an AI detector prove someone used AI?
No. Even Turnitin says its tool “does not make a determination of misconduct” [9], and OpenAI withdrew its own detector for low accuracy [11].
The verdict
The marketing says 99%; the independent evidence says far less, with real risks for non-native writers and anyone whose text gets lightly edited. The decision: use a detector to start a conversation, never to end one.
Related: Is Grammarly’s AI detector accurate? · Do AI humanizers work?
Change log
15 Oct 2026 — Published as an evidence review. Hyptor Lab hands-on test scheduled.
Sources
- Weber-Wulff D. et al., “Testing of detection tools for AI-generated text”, Int J Educ Integrity, 2023 — link.springer.com
- Perkins M. et al., “Simple techniques to bypass GenAI text detectors”, IJETHE, 9 Sep 2024 — springeropen.com
- GPTZero homepage/FAQ, accessed 14 Oct 2026 — gptzero.me
- ZeroGPT homepage, accessed 14 Oct 2026 — zerogpt.com
- Grammarly AI Detector, accessed 14 Oct 2026 — grammarly.com
- QuillBot AI Detector, accessed 14 Oct 2026 — quillbot.com
- Copyleaks AI Detector, accessed 14 Oct 2026 — copyleaks.com
- Scribbr AI Detector, accessed 14 Oct 2026 — scribbr.com
- Turnitin Originality, accessed 14 Oct 2026 — turnitin.com
- Liang W. et al., “GPT detectors are biased against non-native English writers”, Patterns, 10 Jul 2023 — pmc.ncbi.nlm.nih.gov
- OpenAI, “New AI classifier for indicating AI-written text” (updated 20 Jul 2023) — openai.com
- Vanderbilt University, “Why We’re Disabling Turnitin’s AI Detector”, 16 Aug 2023 — vanderbilt.edu
- University of Waterloo, “Discontinuing use of AI detection functionality in Turnitin”, Sept 2025 — uwaterloo.ca
- Dugan L. et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors”, ACL 2024 — arxiv.org
- RAID leaderboard, accessed 15 Oct 2026 — raid-bench.xyz
- Chechitelli A., “AI writing detection update from Turnitin’s Chief Product Officer”, 23 May 2023 — turnitin.com
- Pricing pages (GPTZero, ZeroGPT, Grammarly, QuillBot, Copyleaks), accessed 14 Oct 2026 — gptzero.me, zerogpt.com, grammarly.com, quillbot.com, copyleaks.com
Right of reply
Vendors named here can respond via our contact page; substantive responses will be published here.