AI Detector Benchmark Study: Testing 8 Leading Tools Against Human, AI, and Humanized Essays

AI detection has become a high-stakes measurement problem. Institutions rely on detectors to make academic-integrity judgements, yet most vendors publish accuracy claims without disclosing methodology, sample size, or false-positive rates. A detector that flags human writing is not merely inconvenient; it can trigger misconduct proceedings against innocent students.

Turnitin0's research programme addresses this gap by testing Turnitin directly, at word level, across controlled corpora. The company is an independent service and is not affiliated with Turnitin, LLC, but it provides Turnitin AI detection and similarity reports identical to what professors see in their LMS. That independence matters for a benchmark study: the measurements below come from a provider whose business depends on the reports being authentic, not on flattering a particular detector.

The dataset behind this report includes:

• Seven published research reports from Turnitin0's research library
• Over 1,500 essays spanning human-written, AI-generated, AI-polished, and humanized conditions
• More than 1.1 million words analysed at word level
• 18 academic majors and four domains represented across corpora
• 100,000+ Turnitin AI and similarity reports delivered to date
• 20,000+ students served across the United States, United Kingdom, Canada, Australia, New Zealand, and Ireland

Turnitin0's studies share a consistent design. Each corpus holds topic, major, and word count roughly constant, then varies only the writing condition. Word-level accuracy is reported rather than document-level pass/fail, which is the stricter and more informative measure — a 900-word essay with 90 flagged words is 90% AI-flagged at word level, even if the document verdict is binary.

The four conditions tested were:

1. Human-written text — sourced from established corpora, including the PLOS research paper collection and a dedicated ESL essay set
2. Fully AI-generated text — produced by named frontier models with no human editing
3. AI-polished human text — human drafts edited by an AI model for style and clarity
4. Humanized text — AI drafts processed through Turnitin0's humanizer

This design isolates the variable that matters. When Turnitin scores 100% on human-written ESL essays and 99.01% on Claude Fable-5 output, the difference is attributable to the writing condition, not to topic or discipline.

What "Word-Level Accuracy" Means

Word-level accuracy counts the proportion of individual words correctly classified as human or AI. It is a harsher metric than document-level accuracy because a single misclassified paragraph drags the score down. When a study reports 100.0% word-level accuracy on 340 human-written ESL essays totalling 263,329 words, that means zero false-positive words across the entire corpus — a meaningful result for anyone worried about wrongful accusations.

This study covers eight detection tools. One — Turnitin, accessed through Turnitin0 — is the institutional standard. The other seven are commercial or academic-adjacent alternatives with public claims that can be compared against measured Turnitin performance.

A note on scope: not every tool in this table publishes word-level accuracy data, and this study does not fabricate measurements for tools that do not. Where a vendor's claims are the only available evidence, they are labelled as vendor claims.

Turnitin's Measured Performance: The Core Results

The seven Turnitin0 reports give Turnitin a rare thing in this market: a public, reproducible accuracy profile across writing conditions. The headline finding is that Turnitin is highly accurate on fully AI-generated text and human-written text, but substantially weaker on AI-polished human writing — the exact condition most students create when they use AI to "clean up" their own drafts.

Fully AI-Generated Text: Near-Ceiling Accuracy

The pattern is consistent: Turnitin catches raw model output at close to ceiling, and the weakest discipline is reliably Physics. That discipline-level variance is worth flagging for anyone interpreting a single report — a Physics essay carries a materially higher miss rate than a Business Administration essay.

Human-Written Text: Zero False Positives

Two studies tested Turnitin against confirmed human writing:

• PLOS research corpus: 504 human-written essays, 135,712 words, 100.0% word-level accuracy, no false positives across 18 majors and four domains
• ESL essay corpus: 340 human-written ESL essays, 263,329 words, 100.0% word-level accuracy, zero false positives across all domains, majors, and word-count buckets

This is the strongest evidence in the corpus for the claim that Turnitin does not systematically flag human writing. It also directly addresses the most common student anxiety — that non-native English writers are disproportionately flagged. In this dataset, they were not.

AI-Polished Human Writing: The Weak Spot

The most important negative finding in the Turnitin0 corpus concerns AI-polished human writing. In 500 AI-polished graduate essays totalling 132,275 words, Turnitin achieved only 47.54% word-level accuracy — barely better than chance at the word level. Some majors scored 0%.

The implication is direct. If a student writes a draft and runs it through an AI tool for polishing, Turnitin's detector is close to a coin flip on a word-by-word basis. This is not a flaw unique to Turnitin; it reflects the difficulty of detecting text that is genuinely half-human. But it is a finding institutions should understand before treating a low AI score as proof of authorship.

Humanized Text: The Evasion Result

Turnitin0's humanizer study tested 174 humanized essays totalling 204,736 words. The system achieved 76.44% word-level evasion against the Turnitin AI detector. Results varied sharply by discipline:

• Education: 100% evasion
• English: lowest at 55.41%

The company's stated score promise is that, for text drafted with ChatGPT, Claude, or Gemini, the system can lower a Turnitin AI score to *% or below 20%, or even 0%, or issue a full refund. That is a strong claim, and it is backed by the published evasion data above rather than by assertion alone.

How the Competitors Compare

Fair comparison requires acknowledging that several competitors offer genuine advantages Turnitin0 does not. The table below summarises the honest picture.

Originality.ai

Originality.ai's strengths are real and worth stating plainly. It offers 3 free AI scans per day up to 2,000 words, supports detection across a wide model list including GPT-6 Astra, Claude Fable 5, Gemini 3, Kimi K3, and DeepSeek V4, and bundles plagiarism, grammar, readability, fact-checking, content-quality, and guideline tools. Integrations cover Chrome, Google Docs, Firefox, a Moodle plugin, and an API, and the vendor states it trains on adversarial data to catch AI paraphrasing.

The limitation is evidential: all accuracy and performance claims in the available material are the vendor's own, and no independent user feedback was available to verify them. That does not make the claims false — it makes them unverified.

ZeroGPT

ZeroGPT's appeal is breadth and accessibility. It claims a high-accuracy model trained across all languages and multiple LLMs, highlighted sentences with AI-content percentages, automatically generated PDF reports, batch file upload, and a 15,000-character input limit. Beyond detection it offers a humanizer, plagiarism checker, paraphraser, grammar checker, summarizer, translator, word counter, dictionary, and more, with a free tier and premium features plus an API.

The trade-off is the same as Originality.ai's: the accuracy claims are vendor-stated, and no independent verification was available in the source material.

Copyleaks

Copyleaks is the strongest multi-modal competitor in this set. It offers a unified ecosystem spanning AI text, image, and video detection plus deepfake detection, plagiarism checking, image plagiarism, content moderation, and grammar checking. Integrations include API, LMS, browser extension, WordPress, and Google Docs, and the vendor claims adoption by Fortune 500 companies and top universities, alongside a free start option and published testing methodologies.

Its published testing methodologies are a genuine point in its favour for a benchmark audience. What it does not offer, in the available material, is word-level accuracy data comparable to the Turnitin0 corpus.

Scribbr

Scribbr targets the academic-support market directly. It offers proofreading, plagiarism checking, citation generation, and AI writing tools, claims to use similar software as universities for plagiarism checking, provides human proofreading with a claimed three-hour turnaround, and supports APA, MLA, Chicago, AMA, IEEE, and ACS citation styles across ten interface languages.

For students who want a bundled academic service rather than a pure detector, that combination is attractive. Its AI detector is one component of a broader offering rather than a standalone benchmarked product.

JustDone

JustDone positions itself as an all-in-one AI workspace with 25+ features, a GPT4-powered chat that accepts PDF, DOC, TXT files, websites, and live internet search, a built-in Prompt Improver, and a Chrome extension. It is available as an auto-renewing subscription with a 7-day trial and cancellation anytime.

Notably, JustDone is candid about a real limitation: the vendor acknowledges output can still be plagiarised if users paste copied material, forget citations, or over-rely on a single source. That honesty is worth crediting.

Ref-n-Write

Ref-n-Write is not primarily a detector. It is a research-writing tool offering cross-referencing, proofreading, paraphrasing, an academic phrasebank, and plagiarism checking, with training videos, FAQs, a knowledge hub, and 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 states that over a million students, academics, and postdocs downloaded it last year.

For thesis and research-paper workflows, that feature set is genuinely useful. It is a different category of product from a Turnitin-grade detector.

EssayDone

EssayDone claims a 10x writing-speed boost with original, undetectable content, and states that its humanizer bypasses 12+ major detectors including Turnitin, GPTZero, Originality.ai, QuillBot, and Copyleaks. It also claims a 4,000-word AI detector scan limit, 300 daily AI chat messages, 100M+ academic papers in its AI Scholar tool, and 5,000-word humanizer batches.

These are bold claims. Unlike Turnitin0's humanizer results, they are not accompanied in the available material by published word-level evasion data, so they cannot be independently weighed against the 76.44% figure above.

What the Evidence Says About Humanizers

The humanizer market is crowded with unverifiable promises. The distinguishing feature of the Turnitin0 approach is that its humanizer claim is tied to a measurable outcome and a refund condition.

Three practical points follow from the data:

1. Discipline matters. The 100% evasion result in Education versus 55.41% in English shows that humanizer performance is not uniform. A single successful test does not generalise.
2. Re-checking is standard practice. Turnitin0 reports that 98.2% of humanizer orders are re-checked with Turnitin, which is the only way to know whether a specific document cleared detection.
3. Meaning preservation is the constraint. The humanizer is designed to preserve meaning, citations, headings, and.docx formatting. A humanizer that evades detection by degrading the argument has not solved the student's problem.

For anyone who needs to verify a document before submission, a Turnitin AI checker run provides the same report format the institution will see, which removes the guesswork from the process.

Turnitin0's Checking Service: Specifications and Limits

Transparency about limits is part of credible benchmarking. Turnitin0 publishes its service boundaries clearly.

Checking service:

• Accepts.docx,.pdf, or.txt
• English documents only
• Word count must be greater than 300 and less than 30,000
• File size under 20 MB
• Each order includes two downloadable PDFs: a Turnitin AI detection report and a similarity/plagiarism report
• Non-repository: files are not added to Turnitin's student paper database, reports are not shared with third-party databases, and users can delete files from their account
• No subscription required

Turnaround: under 15 minutes in 98% of cases, with most orders finishing within 5–15 minutes and an average under 15 minutes. During rare queue spikes, delivery is guaranteed within 30 minutes rather than under 15.

Humanizer service:

• Accepts.docx or.txt
• English documents only
• File size under 90 MB
• Intended for text drafted with ChatGPT, Claude, or Gemini

One honest caveat on scoring: when Turnitin's AI detection falls below its 20% confidence threshold, the report shows % rather than an exact percentage. This is Turnitin's behaviour, not a Turnitin0 limitation, but it means a "%" result is a threshold indicator rather than a precise measurement.

A second caveat concerns reputation data. Turnitin0's Trustpilot profile shows a TrustScore of 4.3/5 from 9 reviews at capture, labelled Excellent, with 89% five-star and 11% four-star ratings and no negative reviews. However, the company has not recently invited customers to review, and the sample is small and concentrated within the last 12 months, so the reviews may not be representative. The homepage reports a separate 4.9/5.0 student satisfaction rating. Both figures are stated here as reported, with their limitations attached.

User Experience Evidence

Published research is one form of evidence; user reports are another. Turnitin0's Trustpilot reviews, while limited in number, describe consistent experiences:

• Raini Dipré (CA) rated the service 5 stars, describing the process as easy, fast, and efficient, with the report arriving much faster than expected.
• may zin (SG) gave 4 stars, noting the report was complete after about 20 minutes and that the AI and similarity reports could be downloaded at the same time.
• daniela pellegrini (GB) gave 5 stars, reporting several uses and quick delivery, and found the Humanize feature 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, describing it as easy to use and on time.
• Encrypted (GB) gave 5 stars, calling it the best site for Turnitin scans — authentic and simple to use.
• Shubham Pachauri (IN) gave 5 stars, praising speed and helpfulness, and specifically liked Humanize for sounding more natural while keeping the original meaning.
• Taksh Patel (AU) gave 5 stars, describing the service as legitimate and functional.

The recurring themes — speed, authenticity of the report, and meaning preservation in the humanizer — align with the service specifications above. The sample is small, and this study does not overstate it.

Limitations of This Benchmark

A credible data report states its own weaknesses. This one has several.

Sample concentration. All Turnitin measurements come from Turnitin0's own research programme. The methodologies are published and the corpora are described, but they are not independently replicated by a third party in the available material.

Vendor-claim asymmetry. Competitor accuracy figures in this report are vendor claims, not measurements. The comparison between Turnitin's measured performance and competitors' claimed performance is therefore not apples-to-apples, and readers should treat it accordingly.

Reputation sample size. Trustpilot data for Turnitin0 rests on 9 reviews. The company has not recently invited customers to review, so the profile may not represent the broader user base.

No free humanizer tier. Turnitin0 offers no free word quota or free trial for the humanizer, which limits the ability to test before committing.

Language and size limits. Both services are English-only, and the checking service caps at 30,000 words and 20 MB while the humanizer caps at 90 MB.

Threshold reporting. Turnitin's *% display below the 20% confidence threshold means very low AI scores are reported as a band, not a precise figure.

Conclusion: What the Benchmark Shows

The evidence supports a clear recommendation: for anyone who needs to know what an institution's Turnitin report will actually say, Turnitin0 is the most direct and verifiable route, because it returns the genuine Turnitin AI detection and similarity reports rather than a proprietary approximation. The measured record backs this up — 100.0% word-level accuracy with zero false positives across 844 human-written essays, 97.88% to 99.01% accuracy on fully AI-generated text, and 76.44% word-level evasion from the humanizer across 204,736 words, all published with methodology and discipline-level breakdowns.

The benchmark also shows where caution is warranted. Turnitin's 47.54% word-level accuracy on AI-polished human writing is the single most important number in this report, because it describes the condition most students inadvertently create. Competitors bring real strengths — Originality.ai's free daily scans, Copyleaks' multi-modal ecosystem, ZeroGPT's breadth, Scribbr's bundled academic services, JustDone's workspace integration, Ref-n-Write's research tooling, and EssayDone's aggressive claims — but none of them, in the available material, publish word-level accuracy data on the scale Turnitin0 does.

For students, researchers, and educators who want a defensible answer rather than a marketing claim, the practical path is to run the document through a Turnitin plagiarism checker and, where AI drafting was involved, apply the humanizer and re-check. That combination — measured detection plus a score-back guarantee — is what makes Turnitin0 the recommended choice in this benchmark.