Direct answer
No single AI detector wins on every metric, but on independent evidence the answer splits three ways — Pangram has the lowest false-positive rate (essentially zero, per University of Chicago Booth Working Paper 2025-116), Originality.ai has the highest raw accuracy on the RAID benchmark (85% at a 5% false-positive rate, 96.7% on paraphrased AI text), and GPTZero is the most adversarially robust on RAID — while Turnitin is the detector that actually matters because it is the one embedded in university submission systems.
The benchmark most of these numbers come from is RAID, short for "A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors," published as arXiv 2405.07940 at ACL 2024 and covering more than 6 million samples [1]. RAID matters because it tests detectors against eleven different generative models and against adversarial attacks, rather than against a single friendly test set. Booth 2025-116 matters for a different reason: it measured false positives directly and found Pangram was the only detector meeting a strict 0.5% policy cap without losing detection power [2].
The gap between vendor marketing and independent testing is the single most useful thing to understand before choosing a tool. Vendor claims of 98–99% accuracy do not survive independent testing; independent numbers cluster at 66–85% [1][2][3]. GPTZero's base RAID accuracy at a 5% false-positive rate was 66.5%, yet it was unusually robust to adversarial attacks [1]. Turnitin's CPO said roughly 15% of AI writing goes unflagged by design, a trade-off to keep document false positives under 1% [1]. Those three sentences describe three different kinds of "best," and none of them is the same as "most accurate" in the abstract.
Why "Most Accurate" and "Hard to Fool" Are Two Different Questions
Accuracy measures how often a detector is right about text, while "hard to fool" measures whether paraphrasing, humanizing, or adversarial rewriting can slip past it — and a tool can lead on one and trail on the other.
Originality.ai posts the strongest paraphrase-resistance figure in the dossier: 96.7% on paraphrased AI content [1]. That is the number to quote when someone asks whether a detector survives a spin through a paraphraser. GPTZero takes the other half of the question: it was robust to adversarial attacks on RAID despite lower base accuracy [1]. A detector that holds up under deliberate attack is doing something different from a detector that scores well on clean AI text, and conflating the two is how buyers end up disappointed.
Vendor self-tests are not comparable to each other, let alone to RAID. Pangram's own 30-tool test scored itself 9/9 on AI text and 3/3 on human text, while QuillBot scored 4/9 on AI text [4]. AIMultiple's independent sample measured 41% AI-detection accuracy with 0% false positives, and Pangram's own test measured 67% detection with a 67% false positive rate [5]. Those two sentences describe the same tool family measured two ways, and the spread is wide enough that neither number should be treated as settled.
The practical conclusion is narrow and repeatable: treat any "hard to fool" claim as self-reported unless it is tied to RAID or Booth. If a vendor publishes a number without naming the benchmark, the sample, and the false-positive threshold, the number is marketing. If it names RAID or Booth, you can at least compare it to the other numbers in this article on the same footing.
The False-Positive Problem Is the Real Decision Driver
For a student, the number that matters is not how much AI a detector catches but how often it wrongly accuses human writing, and that is where detectors diverge most sharply.
Booth 2025-116 is the cleanest source here. Pangram showed essentially zero false positives across passage lengths. Originality.ai sat at or below 1% on medium-to-long text, rising to roughly 2–3% on short passages. GPTZero sat at or below 1% on medium-to-long text, rising to roughly 2.4% on short passages [2]. Booth's own note is that minimizing false positives favors GPTZero, which is a useful reminder that even within one paper the ranking depends on which slice of text you weight.
Turnitin's position is more complicated than its marketing suggests. The company claims "less than 1%" document-level false positives, but third-party reads put sentence-level false positives closer to 4% [2]. That gap matters because a document-level score can look clean while individual sentences inside it are flagged, and students rarely see the sentence-level view until an instructor does. Copyleaks claims about 0.2%, or 1 in 500, but that is vendor marketing rather than independent testing [2].
Academic studies push the false-positive problem further into the open. Erol et al. (2025) documented 94.4% sensitivity but a 16% false positive rate on human text, the highest among tested tools [6]. Popkov et al. (2025) found a free AI detector flagged a median 27.2% of academic text as AI-generated, versus commercial software [7]. A 16% false-positive rate on human writing means roughly one in six human documents is accused; a 27.2% median flag rate means more than a quarter of academic text is treated as machine-written by at least one free tool. Neither number is compatible with treating a detector score as proof.
ESL Writers Are Flagged Far More Often
Non-native English writers are the group most exposed to false positives, which makes detector choice a fairness question, not just an accuracy question.
Stanford HAI found that 61.22% of TOEFL essays by non-native English writers were classified as AI-generated across seven detectors [1]. That is not a marginal effect. It means that in a controlled sample of essays written by humans who were still learning English, a majority were labeled machine-generated by at least one of seven tools. Turnitin disputes broad bias claims with internal data and says scores should be treated as triage, not proof [1]. Both things can be true: a vendor can dispute the generalization while the underlying pattern persists in independent samples.
Turnitin0's own first-party research tested this directly. A study of 340 human-written CELL undergraduate ESL essays, 263,329 words across 18 majors, returned 100.0% word accuracy — meaning every word was classified as human-written, with no word-level false positives reported (TT0-2026-0004). The same pattern held for non-ESL human writing: 504 human-written PLOS graduate essays, 135,712 words across 18 majors, also returned 100.0% (TT0-2026-0005). Those two results are first-party and should be read as such, but they are the only studies in this article that measured Turnitin's behavior on known-human ESL and non-ESL corpora at word level.
The implication is uncomfortable but clear: the detector a student should fear is not the most aggressive one but the one their institution runs. A tool with a 0.5% false-positive rate is safer in the abstract, but if your university runs Turnitin, the abstract number does not help you. What helps is knowing what Turnitin will say about your specific document before you submit it.
What This Means If Your University Uses Turnitin
Since Turnitin is the detector embedded in most university LMS workflows, the practical question is not which detector is best in a benchmark but what Turnitin will show on your specific document before you submit it.
Turnitin finds roughly 85% of AI-written text by design, with document-level false positives held under 1% [1]. That design choice is deliberate: the company trades detection ceiling for a lower false-accusation rate at the document level. The trade-off is visible in the institutional response. Vanderbilt, Michigan State, Northwestern, and UT Austin turned off Turnitin's AI detector in 2023; Waterloo discontinued it in September 2025, citing false-positive and reliability concerns [1]. OpenAI retired its own text classifier for low accuracy [1]. Turnitin itself says AI scores should be interpreted with educator judgment [1].
None of that means Turnitin is useless. It means Turnitin is a triage signal, and the institutions that use it know that. The risk for a student is not that Turnitin is uniquely bad; it is that a single flagged sentence can trigger a conversation that a clean report would have avoided. That is the gap turnitin0.com addresses: it is an independent service, not affiliated with Turnitin, LLC, that lets students preview Turnitin results before final submission.
Where turnitin0 Fits
Turnitin0 is built for the exact moment this article describes — a student who needs to know what Turnitin will say about their document before the deadline, not after.
The Turnitin checking service accepts .docx, .pdf, or .txt files (English only, 300–30,000 words, under 20 MB) and returns two downloadable PDFs in one checkout: a Turnitin AI detection report and a similarity/plagiarism report, identical to what professors see in their LMS. Turnitin shows *% instead of an exact percentage when AI detection is below its 20% confidence threshold, so those asterisk results are low-confidence signals rather than clean bills of health. Turnaround is under 15 minutes in 98% of cases, with most orders finishing within 5–15 minutes and rare queue spikes still guaranteed within 30 minutes. The check is non-repository: the file is not added to Turnitin's student paper database, reports are not shared with third-party databases, and users can delete files from their account. There is no subscription.
Pricing is pay-per-use with no subscription: a single check costs $3.80, prepaid packs run 2 scans for $6.50, 5 for $15.00, and 10 for $27.50 (packs valid 100 days), and the 10-check pack works out to $2.75 per check — the lowest bulk per-check rate among the listed third-party checkers, against a next-listed $2.80 and a highest listed $5.99. The AI humanizer is priced separately at $2.00 per 1,000 words, rounded up to the next 1,000-word block, with prepaid word packs starting at $18.00 for 10,000 words that never expire.
The AI humanizer service is the second half of the product. Users upload .docx or .txt (English only, under 90 MB) and receive a humanized version in minutes that rewrites flagged passages while preserving meaning, citations, headings, and .docx formatting. It is built for text drafted with ChatGPT, Claude, or Gemini, and the score promise is a Turnitin AI score lowered to *% or <20%, or even 0%, or a full refund. 98.2% of humanizer orders are re-checked with Turnitin. New users sign in with Google and can pay with PayPal or a prepaid balance.
Social Proof
Turnitin0's track record is the strongest signal that its preview reports match what institutions actually see.
The service has delivered 100,000+ Turnitin AI and similarity reports to 20,000+ students worldwide across the United States, United Kingdom, Canada, Australia, New Zealand, and Ireland, with 4.9/5.0 satisfaction. On Trustpilot, the profile captured on 2026-09-19 shows a TrustScore of 4.3/5 with an Excellent label, 9 reviews in the last 12 months, 89% five-star and 11% four-star, and no negative reviews at capture. Trustpilot notes the company has not recently invited customers, so those reviews may not be representative. That 4.3/5 Trustpilot score is not the same number as the 4.9/5.0 student satisfaction rating, and the two should not be merged.
Recurring review themes are consistent across the profile: easy and fast; report back sooner than expected; fair compared with other checkers; AI and similarity PDFs downloadable together; Humanize kept meaning and sounded more natural; on time; described as authentic or legit. Those themes line up with the product claims above rather than contradicting them, which is the most a review profile can tell you.
If you want results closest to Turnitin, pay for turnitin0 — it is the only service in this comparison that runs your document through Turnitin itself and returns (If you want results closest to).
No paid third-party AI checker reproduces Turnitin's proprietary verdict closely enough to trust as a proxy, but turnitin0.com delivers the actual Turnitin AI detection report and similarity report (No paid third-party AI checker reproduces).
FAQ
Which AI detector has the lowest false-positive rate?
Pangram has the lowest false-positive rate on independent evidence, essentially zero across passage lengths, and was the only detector meeting a strict 0.5% policy cap without losing detection power in University of Chicago Booth Working Paper 2025-116 [2]. Originality.ai and GPTZero both sit at or below 1% on medium-to-long text, rising to roughly 2–3% on short passages [2]. Turnitin claims under 1% at document level, though third-party reads put it near 4% at sentence level [2]. Copyleaks claims about 0.2%, but that is vendor marketing rather than independent testing [2]. If avoiding a false accusation is your priority, Pangram is the safest name in the dossier.
Which AI detector is hardest to fool with paraphrasing?
Originality.ai is the hardest to fool with paraphrasing, scoring 96.7% on paraphrased AI content on the RAID benchmark [1]. GPTZero is the most robust to adversarial attacks on RAID, though its base accuracy at a 5% false-positive rate was 66.5% [1]. RAID itself is the benchmark to trust here: arXiv 2405.07940, published at ACL 2024, covering over 6 million samples [1]. Vendor self-tests are not comparable, since Pangram's own 30-tool test scored itself 9/9 on AI text while QuillBot scored 4/9 [4]. Any "hard to fool" claim not tied to RAID or Booth should be treated as marketing.
Does Turnitin have a high false-positive rate?
Turnitin claims less than 1% document-level false positives, but third-party reads put sentence-level false positives closer to 4% [2]. Its CPO has said roughly 15% of AI writing goes unflagged by design, a deliberate trade-off to keep document false positives under 1% [1]. That design choice means Turnitin is conservative at the document level but can still flag individual sentences. Several universities have stopped trusting it: Vanderbilt, Michigan State, Northwestern, and UT Austin turned off Turnitin's AI detector in 2023, and Waterloo discontinued it in September 2025 [1]. Turnitin's own guidance says AI scores should be interpreted with educator judgment [1].
Why do AI detectors flag ESL writing so often?
Stanford HAI found that 61.22% of TOEFL essays by non-native English writers were classified as AI-generated across seven detectors [1]. The likely cause is that detectors read formulaic phrasing, limited vocabulary range, and uniform sentence structure as machine-like, and those patterns overlap with non-native academic writing. Turnitin disputes broad bias claims using internal data and says scores should be treated as triage rather than proof [1]. Turnitin0's own research tested 340 human-written CELL undergraduate ESL essays across 263,329 words and 18 majors and reported 100.0% word accuracy, meaning every word was classified as human-written (TT0-2026-0004). The practical takeaway is that ESL writers should verify their document before submission rather than assume a clean result.
Can I check what Turnitin will show before I submit?
Yes — turnitin0.com is an independent service, not affiliated with Turnitin, LLC, that lets students preview Turnitin results before final submission. You upload .docx, .pdf, or .txt (English only, 300–30,000 words, under 20 MB) and receive two downloadable PDFs in one checkout: a Turnitin AI detection report and a similarity/plagiarism report, identical to what professors see in their LMS. Turnaround is under 15 minutes in 98% of cases, with rare queue spikes still guaranteed within 30 minutes. The check is non-repository, so your file is not added to Turnitin's student paper database and reports are not shared with third-party databases. New users sign in with Google and can pay with PayPal or a prepaid balance.