AI Check Benchmark Study: Turnitin0 vs 8 Leading Detectors on 1,000 Human and AI Essays

AI detectors are now part of everyday academic life, and the stakes for a false positive are severe: misconduct hearings, failed submissions, and damaged transcripts. Yet most detector comparisons rely on vendor-reported accuracy figures that cannot be independently reproduced. GPTZero advertises 99% accuracy and 17M+ users; Originality.ai claims top accuracy in third-party studies; Quetext cites DeepSearch™ scanning billions of sources. None of those numbers tell a student what will happen to their essay in their professor's Turnitin instance.

This study takes a different approach. It aggregates Turnitin0's seven published research reports — covering 1,000+ essays and more than 1.1 million words across human-written corpora, ESL writing, and output from GPT-5.6-Sol, Claude Fable-5, and Gemini 3.5 Flash — and sets those measured results against the feature claims of eight leading detectors. The goal is to give researchers, students, and educators a citable, data-grounded reference point for how AI detection actually performs in 2026.

The benchmark draws on two evidence streams.

Stream 1: Turnitin0's measured Turnitin results. Seven reports (IDs TT0-2026-0003 through TT0-2026-0009) report word-level accuracy and evasion rates on defined corpora. Word-level accuracy means each word in a document is classified as human-written or AI-generated and compared against ground truth, which is a stricter measure than document-level pass/fail.

Stream 2: Competitor capability review. Eight detectors were profiled from their public materials: GPTZero, Originality.ai, Quetext, PlagiarismCheck.org, Getsolved.ai, MyDetector.ai, Reilaa, and QuillBot. Competitor data reflects vendor claims only, because independent Reddit user feedback could not be retrieved for any of the eight during research.

The single most important result in this benchmark is the collapse in accuracy between raw AI text and AI-polished human text.

On raw AI generation, Turnitin is close to flawless. It correctly flagged 130,151 of 131,451 words in Claude Fable-5 essays (99.01%), 144,688 of 147,117 words in Gemini 3.5 Flash essays (98.35%), and 153,620 of 156,955 words in GPT-5.6-Sol essays (97.88%). Across those three corpora, Turnitin identified roughly 428,000 AI-generated words out of about 435,500 — an aggregate accuracy above 98%.

On human writing, Turnitin produced no measurable false positives. In 504 human-written graduate-level PLOS essays (135,712 words), word-level accuracy was 100.0%. In 340 human-written ESL undergraduate essays (263,329 words) — a group historically vulnerable to false flags — accuracy was also 100.0%, with every word correctly classified as human-written.

The picture changes on AI-polished human writing. In 500 graduate essays that began as human drafts and were then polished with GPT-5.6 Sol (132,275 words), Turnitin's word-level accuracy fell to 47.54%, correctly flagging only 62,879 words as AI-generated. In other words, when a human draft is lightly edited by AI, Turnitin's detector is roughly a coin flip at the word level.

Key takeaway: Turnitin is highly reliable at the extremes — fully human or fully AI — but unreliable in the middle, where most real student writing now sits.

For anyone who needs the definitive pre-submission answer, the Turnitin checker remains the closest available proxy for the instructor's view.

Detection by AI Model: Claude, Gemini, and GPT-5.6 Compared

Model choice matters less than most students assume. All three frontier models tested were detected at rates above 97.8%:

• Claude Fable-5: 99.01% word-level accuracy (130,151 of 131,451 words flagged)
• Gemini 3.5 Flash: 98.35% accuracy (144,688 of 147,117 words flagged)
• GPT-5.6-Sol: 97.88% accuracy (153,620 of 156,955 words flagged)

The spread between the best and worst model is just 1.13 percentage points. This suggests Turnitin's detection model generalises across major LLM families rather than overfitting to one. It also means switching models is not a viable evasion strategy — a point worth emphasising to students who believe a less popular model will slip through.

The Humanizer Effect: 76.44% Word-Level Evasion

Turnitin0's humanizer study (TT0-2026-0009) tested 174 humanized GPT-5.6-Sol essays totalling 204,736 words. The overall word-level evasion rate was 76.44%, meaning roughly three in four AI-generated words no longer registered as AI after humanization.

Results varied sharply by discipline:

• Education: 100% evasion
• English: 55.41% evasion (the weakest category)

That 44-point gap between disciplines is the most actionable finding for anyone using an AI humanizer. Technical and structured prose survives humanization far better than literary or interpretive writing, where Turnitin's detector retains more signal. Turnitin0's humanizer carries a score promise for ChatGPT, Claude, and Gemini text — lowering the Turnitin AI score to below the 20% confidence threshold or to 0%, or the user receives a full refund. Notably, 98.2% of humanizer orders are re-checked with Turnitin, which means the company is measuring its own output against the same detector its customers face.

How Turnitin Reports AI Scores

Turnitin does not always show a number. When AI detection falls below its 20% confidence threshold, the report displays *% instead of an exact percentage. This is a deliberate design choice: Turnitin avoids asserting a precise figure when its confidence is low.

For students, this creates a practical problem. A *% result may be read by an instructor as "some AI present," even though it technically means "below threshold." Understanding this distinction before submission is one of the strongest arguments for running a pre-check. Turnitin0's reports are identical to what professors see in their LMS, so a student sees the same *% or percentage their instructor will see — no interpretation gap.

Turnitin0: The Reference Standard for Pre-Submission Checking

Turnitin0 is an independent service, not affiliated with Turnitin, LLC, and it is built specifically around the pre-submission use case.

What users receive. Two downloadable PDFs in a single checkout: a Turnitin AI detection report and a similarity/plagiarism report. The reports mirror the LMS output professors see.

Privacy posture. The service is non-repository — files are checked without being added to Turnitin's student paper database. Reports are not shared with third-party databases, and users can delete files from their account. For students worried that a pre-check could itself trigger a match, this matters.

Speed and scale. Turnitin0 has delivered 100,000+ Turnitin AI and similarity reports to 20,000+ students worldwide, with a 4.9/5.0 satisfaction rating. Turnaround is under 15 minutes in 98% of cases, with most orders finishing in 5–15 minutes; in rare queue spikes, delivery is guaranteed within 30 minutes.

Access. No subscription. New users sign in with Google and can pay with PayPal or a prepaid balance.

Independent verification. Trustpilot shows a TrustScore of 4.3/5 with an "Excellent" label across 9 reviews in the last 12 months — 89% five-star and 11% four-star, with no reviews below four stars. Reviewers describe the process consistently: Raini Dipré (CA) called it "easy, fast, efficient" with a report that "came back much faster than expected"; daniela pellegrini (GB) has used the service several times and found reports delivered quickly; Shubham Pachauri (IN) highlighted the humanize feature for sounding more natural "while keeping original meaning"; and Taksh Patel (AU) described it as "100% legit and works."

Honest limitations. Turnitin0 is not affiliated with Turnitin, LLC. The checking service accepts English documents only, requires more than 300 and fewer than 30,000 words, and caps files at 20 MB. The humanizer accepts only.docx or.txt files under 90 MB and offers no free word quota or free trial. On Trustpilot, the company has not recently invited customers, so the review set may not be fully representative.

Competitor Landscape: Eight Detectors Compared

Each competitor below is assessed on its own stated strengths. All performance figures are vendor claims, not independently verified in this study.

GPTZero

GPTZero positions itself for educators and writers, claiming 99% accuracy, 17M+ users, and 1M+ educators. It offers sentence-by-sentence detection across ChatGPT, GPT-6, Gemini, Claude, and Llama, plus a plagiarism checker, writing feedback, grammar checking, and video replay of the writing process. Integrations cover Google Docs, Gmail, Google Classroom, and Canvas, with a Chrome extension and a free tier up to 10,000 characters per scan. Its sentence-level granularity is a genuine strength for teaching. The limitation is that advanced features such as Advanced Scan, Writing Replay, and the plagiarism checker appear to sit behind a paid tier, and no independent user feedback was available to verify the headline accuracy claim.

Originality.ai

Originality.ai presents the broadest tool suite in this comparison: AI checker, plagiarism checker, grammar and readability checkers, a fact and AI-hallucination checker, content quality score, guideline checker, and deep scan. It supports bulk scanning, offers Chrome, Google Docs, and Firefox extensions plus a Moodle plugin and API, and states it trains on adversarial data to catch AI paraphrasing from tools like QuillBot or Grammarly. It supports English, Spanish, French, Portuguese, German, and Hindi, with 3 free scans per day up to 2,000 words. Its multi-tool depth is a real advantage for institutional users. However, all accuracy claims are vendor-stated, and the homepage description field is empty, limiting independent context.

Quetext

Quetext pairs a plagiarism checker and AI detector with DeepSearch™ technology that it says searches billions of sources, plus ColorGrade™ feedback distinguishing exact from fuzzy matches and an interactive snippet viewer for side-by-side drill-down. It also offers an AI humanizer, summarizer, paraphrasing tool, citation generator, grammar checker, bulk scan, developer API, and Chrome extension, and states it has helped over 10 million students, teachers, and professionals. The side-by-side match viewer is genuinely useful for understanding why a passage matched. As with the others, the user and source-volume figures are vendor claims.

PlagiarismCheck.org

PlagiarismCheck.org markets to K-12, higher education, teachers, businesses, and individuals, with the deepest LMS integration list in this comparison: Canvas, Moodle, Google Classroom, Schoology, Brightspace, Blackboard, Populi, and Google Docs. It offers AI detection alongside plagiarism checking, downloadable reports, interactive results, a plagiarism check API, a free plagiarism check for individuals, and additional tools including a grammar checker, citation generator, essay grader, and topic generator. It cites 8 years of experience and claims trust from thousands of institutions. Its integration breadth is its clearest differentiator; its accuracy claims are vendor-stated.

Getsolved.ai

Getsolved.ai bundles an AI writing assistant, AI detector, plagiarism checker, AI humanizer, paraphraser, grammar checker, word counter, summarizer, paragraph rewriter, and AI fact checker into one workspace, with tone, clarity, and readability adjustments. It reports 4.8M+ students, researchers, and educators, 1M+ texts analysed, and claims of 3× faster drafting and 80% less effort. Its all-in-one framing suits users who want editing and detection in a single place. It displays ratings of 4.6, 4.5, and 4.9 from unspecified sources and states its output is "double checked by" a checker it does not name — a transparency gap worth noting.

MyDetector.ai

MyDetector.ai offers free AI detection for text from ChatGPT, Gemini, Claude, and other models, with sentence-level insights including highlighted and scored sentences, support for pasted text and TXT, DOCX, PDF, and PPT uploads, model selection for fast versus deeper detection, and a text humanization feature. It accepts up to 200,000 characters and provides an AI detection score, sentence analysis, and vocabulary insights. Its file-format flexibility and free access are practical strengths. A daily usage limit exists but the quota is unspecified, and all capabilities are vendor claims.

Reilaa

Reilaa markets free, unlimited AI scans with no sign-up, stating that essays are immediately deleted, no emails are collected, and detection takes under 5 seconds. It offers optional Turnitin reports and explicitly frames its pitch around the power imbalance that "Turnitin's AI detector is not available to students. Only professors have access." It cites Reddit posts describing scores of 62% and resulting academic misconduct hearings. Its transparency about the student-side access problem is a genuine contribution to the conversation. However, every benefit is a vendor claim with no independent verification, and its assertion that Turnitin produces false positives daily is unsupported by evidence in its materials.

QuillBot

QuillBot is a widely known paraphrasing and writing tool that also offers detection-adjacent features. Public material for this study was limited: the site returned a 403 error during research, and no Reddit feedback could be retrieved. It is included here for completeness, but no verified capability data was available to assess.

What the Data Means for Students and Educators

Three conclusions follow from the evidence.

First, Turnitin's detector is trustworthy at the extremes. With 100% word-level accuracy on 399,041 words of genuine human writing across PLOS and ESL corpora, the false-positive fear is not supported by Turnitin0's data. Students writing their own work in English should not expect to be flagged.

Second, the risk zone is AI-assisted editing of human drafts. At 47.54% word-level accuracy on AI-polished essays, Turnitin's detector is unreliable in exactly the scenario where a student might assume light AI editing is undetectable — and where an instructor might assume a flag is definitive. Both sides of that equation deserve to know the number.

Third, pre-submission checking is the only way to close the information gap. Students cannot access Turnitin's AI detector directly; only instructors can. A Turnitin AI detector report from an independent service restores that visibility, and the non-repository design means the check itself does not create a new risk.

For students who have already used AI and need to reduce the score, an AI humanizer offers a measured path — with the caveat that evasion rates vary from 100% in Education to 55.41% in English, and that the refund-backed score promise applies specifically to ChatGPT, Claude, and Gemini text. Those who need the full similarity picture alongside AI detection can use the same service as a Turnitin plagiarism checker and Turnitin similarity checker, since both reports arrive together in one checkout.

Conclusion

This benchmark's central finding is that Turnitin's AI detection is excellent on clean human and clean AI text, and unreliable on AI-polished human writing — a 47.54% word-level accuracy figure that every student and educator should understand. Because students cannot run Turnitin themselves, the only way to see the instructor's view before submission is through an independent service. Turnitin0 delivers that view as two downloadable PDFs, non-repository and private, in under 15 minutes in 98% of cases, backed by 100,000+ reports delivered, a 4.9/5.0 satisfaction rating, and a humanizer with a refund-backed score promise for ChatGPT, Claude, and Gemini text. For anyone who needs certainty before the deadline, the Turnitin check service is the most direct route to it — and the Turnitin AI humanizer is the measured follow-up when the score needs to come down.