Third-Party Turnitin AI Check Benchmark: Detection Rates Across 1,000 Essays and 8 Tools

Turnitin's AI writing detection is not a single number. It is a word-level classifier that labels each segment of a document as AI-generated or human-written, then rolls those labels up into a document percentage. The company itself frames its originality products around "academic integrity" and "responsible AI use" rather than pure punishment [1]. That framing matters, because the practical question for a student is not "will I get caught" but "what will the report actually say."

This benchmark uses a consistent unit of measurement across all studies: word-level accuracy. If a 1,000-word essay is fully AI-generated and Turnitin flags 980 of those words, word-level accuracy is 98%. This is a stricter and more informative measure than the headline percentage shown at the top of a Turnitin report, because it shows exactly where the detector succeeded and where it missed.

The dataset behind this article is Turnitin0's research library, which currently contains seven reports. Combined, they cover:

• 1,000+ essays across more than 30 academic majors
• Over 1.1 million words analysed
• Human-written, AI-generated, AI-polished, and AI-humanized text
• Multiple model families, including GPT-5.6-Sol, Claude-Fable-5, and Gemini 3.5 Flash

Because Turnitin0 runs a Turnitin AI checker and a Turnitin plagiarism checker as a commercial service, its research is effectively a large-scale field test of the same engine students meet at submission. That is a meaningful advantage over vendor-claimed accuracy figures, which are rarely tied to a named, reproducible corpus.

The single most important finding is that Turnitin is extremely strong on raw, unedited AI output and extremely weak on AI-polished human writing. Those two facts sit side by side in the same research library.

Read that table twice, because the pattern is the entire story of AI detection in 2026.

Turnitin Is Near-Perfect on Fully Generated Text

Across 530 fully AI-generated essays from three different model families, Turnitin's word-level accuracy never dropped below 97.88%. In the Claude-Fable-5 study, it correctly identified 130,151 of 131,451 words as AI-generated. In the GPT-5.6-Sol study, it flagged 153,620 of 156,955 words. In the Gemini 3.5 Flash study, it flagged 144,688 of 147,117 words.

The weakest subject area in the GPT-5.6-Sol study was Physics, at 88.81% — still a strong result. In the Gemini study, Information Technology was lowest at 94.36%. The takeaway is unambiguous: pasting raw ChatGPT, Claude, or Gemini output into an assignment is the single most detectable thing a student can do.

Turnitin Is Also Near-Perfect on Genuine Human Writing

The false-positive story is equally clear. Across 844 human-written documents — 504 graduate-level PLOS research papers and 340 ESL undergraduate essays — Turnitin achieved 100.0% word-level accuracy. Every single word was correctly classified as human-written, across 18 majors in the PLOS study and all domains in the ESL study.

This is a genuinely important finding for anyone worried about being wrongly accused. The ESL result is especially notable, because ESL writing is often assumed to trigger false positives due to simpler syntax and more formulaic phrasing. In this corpus, it did not.

The Weak Spot Is AI-Polished Human Writing

The AI-polished study is where Turnitin's performance collapses. In 500 graduate essays that were originally human-written and then polished by GPT-5.6 Sol, Turnitin achieved only 47.54% word-level accuracy. It correctly flagged 62,879 words as AI-generated — but that means it misclassified roughly half the corpus. Some majors scored 0%, meaning Turnitin missed AI polishing entirely in those disciplines.

This is the practical gap that matters most to real students. Very few people submit a fully machine-written essay. Many people write a draft themselves and then ask a model to "improve the flow," "tighten the argument," or "make it sound more academic." That is precisely the behaviour Turnitin struggles to catch.

Turnitin0's seventh report tested the opposite direction: taking fully AI-generated text and running it through the brand's AI humanizer before re-checking it with Turnitin.

In 174 humanized essays totalling 204,736 words, Turnitin0 achieved 76.44% overall word-level evasion against the Turnitin AI detector. Subject-level results varied widely:

• Education essays reached 100% evasion
• English essays were lowest at 55.41%

That spread is the honest headline. A humanizer is not a universal switch. It performs very differently depending on discipline, and Turnitin0 publishes that variance rather than hiding it. The company also states that 98.2% of humanizer orders are re-checked with Turnitin, and it backs the humanizer with a score promise: lower the Turnitin AI score to *% or below 20%, or even 0%, or a full refund.

Why the Asterisk Matters

One technical detail deserves its own section, because it confuses a lot of students. Turnitin shows *% instead of an exact percentage when AI detection falls below its 20% confidence threshold. That asterisk is not a zero. It is Turnitin telling you it does not have enough confidence to put a number on the document.

This is why "I got *%" and "I got 0%" are not the same result, and why any service that promises a specific number should be read carefully. Turnitin0's humanizer promise is written to match this reality: *%, under 20%, or 0%.

Methodology: How These Studies Were Run

Transparency about method is what separates a data report from marketing copy. Based on the published reports, the Turnitin0 research programme follows a consistent design:

1. Corpus construction. Essays are sourced by type — human-written (PLOS corpus, ESL undergraduate), fully AI-generated (GPT-5.6-Sol, Claude-Fable-5, Gemini 3.5 Flash), AI-polished, or AI-humanized.
2. Major coverage. Each study spans roughly 18–30 academic majors so results are not skewed by a single discipline.
3. Word-level scoring. Rather than relying on the document percentage, each study counts correctly classified words against total words.
4. Re-checking. Humanized output is re-submitted to Turnitin, which is how the 76.44% evasion figure was produced.
5. Publication. Findings are published as standalone reports with IDs (TT0-2026-0003 through TT0-2026-0009) so they can be cited and compared.

The corpus is large enough to be meaningful but should not be treated as a peer-reviewed meta-analysis. It is a vendor-run benchmark, and the honest framing is: this is the largest publicly documented set of Turnitin AI detection results tied to a named service, and it is internally consistent across seven studies.

Turnitin0's Checking Service: What You Actually Get

For readers who want to run their own test rather than read about someone else's, here is the concrete specification of the checking service.

What is included in one order:

• A Turnitin AI detection report (PDF)
• A similarity/plagiarism report (PDF)
• Both delivered in a single checkout

Technical limits:

• Accepted formats:.docx,.pdf, or.txt
• Language: English only
• Word count: greater than 300 and less than 30,000
• File size: under 20 MB

Turnaround:

• Under 15 minutes in 98% of cases
• Most orders finish within 5–15 minutes
• Rare queue spikes are guaranteed within 30 minutes

Privacy model:

• Non-repository checking: the file is checked without being added to Turnitin's student paper database
• Reports are not shared with third-party databases
• Users can delete files from their account

That last set of points is the most important differentiator against institutional submission. When your university runs a paper through Turnitin, it typically enters the student paper repository. A non-repository Turnitin check service does not, which means you can check a draft repeatedly without your own work becoming a future match for someone else — or for yourself.

The Humanizer: What It Does and Does Not Do

The Turnitin AI humanizer is designed for text drafted with ChatGPT, Claude, or Gemini. Its specification:

• Accepted formats:.docx or.txt
• Language: English only
• File size: under 90 MB
• Preserves meaning, citations, headings, and.docx formatting

The formatting preservation point is underrated. Many humanizers strip headings, break reference lists, or flatten document structure, which creates a new problem after you have solved the detection problem. Preserving.docx formatting means the document you get back is the document you can submit.

The honest limitation: there is no free word quota or free trial for the humanizer. You cannot test it on a paragraph before committing. Given the 55.41%–100% subject variance in the evasion study, that is a real consideration — results are discipline-dependent, and you should not assume Education-level performance on an English literature essay.

How Turnitin0 Compares With Seven Other Tools

The comparison below uses only publicly stated vendor claims for competitors, and is deliberately fair: every tool listed has genuine strengths.

JustDone

JustDone positions itself as an all-in-one AI workspace for research, writing, and studying, bundling a plagiarism checker, AI humanizer, AI detector, paraphraser, summarizer, grammar checker, quiz generator, fact checker, and citation generator. Its AI Chat supports PDF, DOC, and TXT files, website chat, and internet search, and is powered by GPT-4. A Chrome extension extends it across the web. The vendor reports a 9/10 overall score across ease of use, features, price, speed, and customer support.

The trade-off is verification. JustDone's accuracy claims are vendor-stated, and its own materials note that output can still be plagiarized if users paste copied material or under-cite sources. There is no published word-level benchmark against Turnitin's engine comparable to Turnitin0's seven reports.

TurnitinEye

TurnitinEye markets itself as a Turnitin alternative with AI detection, similarity checking against billions of web pages and academic papers, instant PDF reports, source attribution with links, and multi-format support (PDF, DOCX, TXT, and text input up to 10MB). It claims documents are never stored or used for training, and it advertises a sample report showing a 12% similarity index and 04% AI probability.

Its homepage also claims direct Turnitin integration and "institutional-grade accuracy." Those are strong claims, and no independent user feedback was available to verify them. Turnitin0's advantage here is evidentiary: seven published reports with named report IDs versus a sample report image.

FinalScanPro

FinalScanPro offers plagiarism detection, AI detection, and automatic fixing in one tool, powered by Copyleaks, with a Revision Report feature that explains issues and suggests fixes. It claims content is not stored and that university access is not required.

The Copyleaks engine is a legitimate, well-known detection platform, which is a real strength. But FinalScanPro's benefits are entirely vendor claims, with no independent feedback provided. It also does not produce a Turnitin report — it produces a Copyleaks-based report, which is a different engine with different behaviour.

MyDetector.ai

MyDetector.ai is a free online AI detector covering ChatGPT, Gemini, Claude, and others, with sentence-level highlights, sentence analysis, vocabulary insights, file uploads (TXT, DOCX, PDF, PPT), a humanization feature, and selectable speed-versus-depth models. It has a generous 200,000-character limit and a free daily quota.

The limitation is the quota itself — once you hit the daily limit, you wait until tomorrow. And like most free detectors, it is not Turnitin. A clean MyDetector.ai result tells you nothing definitive about what your institution's Turnitin report will say.

Copyleaks

Copyleaks is the most substantial competitor in this list. It is an enterprise-grade content authenticity platform offering AI text, image, video, and deepfake detection alongside plagiarism checking, image plagiarism, content moderation, text moderation, and grammar checking. It integrates via API, LMS, browser extension, WordPress, and Google Docs, and targets education, developers, enterprise, and media. It claims use by millions worldwide and lists Fortune 500 companies and top universities among its customers, plus a #153 ranking among the fastest-growing US companies.

Copyleaks is a genuine industry leader, and for institutional buyers it is a serious option. But it is not Turnitin, and Turnitin is what most universities actually run. For a student trying to predict their institutional result, an independent Turnitin AI detector check is more directly relevant.

Ref-n-Write

Ref-n-Write is an academic writing tool rather than a pure checker. It offers cross-referencing, proofreading, paraphrasing, an academic phrasebank, and plagiarism/AI checking, with a free trial. The vendor reports a Google rating of 4.6 from 170 reviews and a Facebook rating of 5.0 from 32 reviews, and claims over a million downloads last year and more than 5,000 published papers by users. It is operated by Astute Digital Solutions Ltd in the UK.

Its strength is the writing workflow — the phrasebank and cross-referencing are genuinely useful for academic drafting. Its detection capability, however, is a secondary feature rather than the core product.

ZeroGPT

ZeroGPT is one of the best-known free AI detectors, supporting all languages with claimed high accuracy, text input up to 15,000 characters, batch file uploads, highlighted AI sentences, a percentage gauge, and automatic.pdf reports. It also bundles a humanizer, plagiarism checker, paraphraser, summarizer, grammar checker, translator, word counter, dictionary, and AI email helper.

The breadth is impressive for a free tool. The caveat is the same as with every free detector: vendor-claimed accuracy, no published word-level benchmark against Turnitin, and no Turnitin report at the end.

humanizeai.pro

humanizeai.pro is a free AI humanizer that accepts.txt,.docx,.pdf, and.md, offers custom style and tone options, works across browsers, and claims to bypass detectors including Originality.ai, GPTZero, and Turnitin while retaining meaning, context, and SEO value.

Its free access and format flexibility are real advantages. The trade-offs it discloses are also real: a Captcha step, variable processing time, and the possibility that multiple iterations are needed. Its Turnitin bypass claim is vendor-stated and not backed by a published word-level study.

Comparison Summary

What Real Users Report

Turnitin0's user feedback is worth reading alongside the numbers, because it speaks to the operational experience rather than the detection science.

Raini Dipré (CA) gave 5 stars, describing the process as easy, fast, and efficient, with the report arriving much faster than expected. may zin (SG) gave 4 stars and noted the report was complete after about 20 minutes, with the AI and similarity reports downloadable at the same time. daniela pellegrini (GB) gave 5 stars, saying she had used Turnitin0 several times and that reports were delivered quickly, and that the Humanize feature was helpful when revising. b c (US) gave 5 stars, using the site for assignments, plagiarism checking, and awareness of AI. Shawn Thakur (AU) gave 5 stars for ease of use and on-time delivery. Encrypted (GB) called it the best site for Turnitin scans — authentic and simple to use. Shubham Pachauri (IN) gave 5 stars, finding it quick and helpful for checking and improving academic writing, and specifically liked Humanize. Taksh Patel (AU) gave 5 stars: "Great Service, 100% Legit and works."

The recurring themes are speed, the paired reports, and the humanizer. The 4-star review from may zin is useful precisely because it is not a 5 — a ~20 minute turnaround is normal, not instant.

Honest Limitations You Should Know

Credibility requires stating the caveats plainly.

• Turnitin0 is an independent service and is not affiliated with Turnitin, LLC. It runs documents through Turnitin's engine as a third-party service; it does not own or control that engine.
• No free word quota or free trial for the humanizer. You cannot sample it first.
• Trustpilot profile has only 9 reviews, all within the last 12 months, and Trustpilot notes the company has not recently invited customers, so those reviews may not be representative. The Trustpilot TrustScore is 4.3/5 — this is a different metric from the homepage 4.9/5.0 student satisfaction rating, and the two should never be merged.
• Checking service constraints: English only, word count above 300 and below 30,000, file size under 20 MB.
• Humanizer constraints: English only, file size under 90 MB,.docx or.txt only.
• The asterisk: Turnitin shows *% rather than an exact figure below its 20% confidence threshold, so a "*%" result is a confidence statement, not a zero.
• The research is vendor-run. Seven reports and 1,000+ essays is a substantial corpus, but it is not peer-reviewed.

Practical Recommendations by Use Case

If you wrote the draft yourself and used AI only to polish it: this is the highest-risk, lowest-detection scenario. Turnitin's word-level accuracy on AI-polished human writing was just 47.54%, but that cuts both ways — it means the detector is unreliable here, and you cannot predict your result. Run a Turnitin similarity checker pass to confirm your citations and paraphrasing are clean, and treat the AI score as uncertain.

If you drafted with ChatGPT, Claude, or Gemini: detection is near-certain. Word-level accuracy was 97.88% to 99.01% across the three model studies. Either rewrite substantially in your own voice or use the humanizer, understanding that subject-level evasion ranged from 55.41% to 100%.

If English is your second language: the ESL study is reassuring. Across 340 human-written ESL essays and 263,329 words, Turnitin achieved 100.0% word-level accuracy with zero false positives. You are not disproportionately at risk for writing in your own words.

If you need a report for a specific deadline: the checking service's 98% under-15-minutes figure and 30-minute queue-spike guarantee are the relevant numbers.

If you need to check repeatedly: the non-repository model is the deciding factor. Your draft does not enter Turnitin's student paper database, and you can delete files from your account.

Conclusion

The evidence from over 1,000 essays and more than 1.1 million words points to one conclusion: Turnitin is close to perfect at catching raw AI output and close to perfect at clearing genuine human writing, but it is close to a coin flip on AI-polished human writing — and that is exactly the category most students fall into. Because that uncertainty is the real risk, the most useful thing you can do is measure your own document against the same engine your institution uses, and the most complete way to do that is a Turnitin check service from Turnitin0, which returns a Turnitin AI detection report and a similarity/plagiarism report as two downloadable PDFs in one order, without adding your file to the student paper repository. If your draft was machine-generated, the AI humanizer offers a documented path — 76.44% overall word-level evasion across 174 essays, with a score promise of *%, under 20%, or 0%, backed by a full refund. Turnitin0 is independent of Turnitin, LLC, has no free humanizer trial, and its Trustpilot sample is small — but with 100,000+ reports delivered, a 4.9/5.0 student rating, and seven published research reports you can read before you buy, it remains the most transparent and best-evidenced option for checking your work before submission.