Turnitin AI Detection Accuracy Test: 1,000 Essays Across 8 Humanizers and Detectors, with Turnitin0 Results

This is an original data report. It aggregates seven Turnitin0 research reports covering 1,548 essays and roughly 1.17 million words, and it compares eight tools in the humanizer and detector market: Turnitin0, Humanize AI (humanizeaitext.ai), T-detector (turnitindetector.com), Phrasly, Winston AI, Getsolved.ai, turnitindetectorai.com, and humanizeai.pro, with QuillBot appearing as a secondary reference point.

The goal is narrow and empirical. Not "does AI detection work in theory," but: what accuracy has actually been measured, where does it break down, and what should a student do about it on a Tuesday night with a deadline on Friday.

Methodology and Data Sources

All Turnitin0 figures below come from published Turnitin0 reports, each of which states its corpus size, word count, and discipline breakdown. Word-level accuracy means the share of individual words in the corpus that Turnitin's AI Writing Report classified correctly. Evasion rate is the inverse for humanized text: the share of words that stopped being flagged after humanization.

Three limitations apply and are stated up front. First, Turnitin's own documentation warns that its AI writing detection model "may not always be accurate" and "should not be used as the sole basis for adverse actions against a student" [1]. Second, Turnitin's AI Writing Report only evaluates qualifying text — prose sentences in a long-form format — and requires at least 300 words to generate a score at all [1][2]. Third, Turnitin0's own checking service is English-only, accepts files between 300 and 30,000 words under 20 MB, and shows any AI score below the 20% confidence threshold as *% rather than an exact figure.

Turnitin launched its AI writing detection feature in April 2023 with a stated 98% confidence rate for detecting AI writing tools, integrated into Feedback Studio, Turnitin Originality, Turnitin Similarity, and LMS integrations [3]. The Turnitin0 corpus tests let us check that claim against real essays rather than vendor language.

Read that table twice, because the story is in the bottom three rows, not the top three.

Raw LLM Output Is Easy to Catch

On unedited model output, Turnitin is close to the 98% confidence figure it advertised. Claude Fable-5 essays were caught at 99.01% word-level accuracy, with Criminal Justice highest at 99.80% and Physics lowest at 96.52%. Gemini 3.5 Flash landed at 98.35%, ranging from 94.36% in Information Technology to 99.82% in Business Administration. GPT-5.6-Sol came in at 97.88%, with Physics again the weakest discipline at 88.81% and Business Administration the strongest at 99.67%.

The discipline pattern repeats across all three models: quantitative and technical writing is harder to flag than business, criminal justice, and education writing. If you are submitting a physics lab report generated by an LLM, the detector is measurably less reliable than it is on a marketing case study.

Human Writing Produces Zero False Positives

Two corpora tested whether Turnitin wrongly flags genuine human writing. Across 504 human-written PLOS research papers spanning 18 majors, Turnitin scored 100.0% word-level accuracy with no false positives. Across 340 human-written ESL essays covering 263,329 words, the same result: 100.0% accuracy, zero false positives across all domains and majors.

This is the strongest single finding in the dataset for anyone worried about being accused unfairly. Turnitin's false-positive rate on authentic human prose in these corpora was zero. The vendor's own caution about misidentification [1] remains valid as a general principle, but it did not materialise in 844 human-written documents.

The weakest result in the entire dataset is the one that matters most to real students. In 500 graduate essays that were human-written and then AI-polished with GPT-5.6 Sol — 132,275 words total — Turnitin achieved just 47.54% word-level accuracy. Some majors scored 0%.

Think about what that means in practice. A student writes a genuine draft, runs it through a model for grammar, flow, and clarity, and submits. Turnitin's AI Writing Report is now roughly a coin flip at the word level. The student may be flagged for text they actually wrote and merely edited with AI assistance, or they may pass cleanly despite substantial AI involvement. Neither outcome is a reliable signal of authorship.

This is the single most important reason to check before submitting rather than after. The Turnitin plagiarism checker and AI detector embedded in your LMS will produce a number, but on polished human writing that number carries far less diagnostic weight than most instructors assume.

Why the Gap Exists

Turnitin's detection model targets patterns associated with generative tools, and it explicitly detects text that "may have been further modified by an AI bypasser, AI-paraphrasing tool or AI word spinner" — though only the English detector includes paraphrasing and bypasser detection, not the Spanish or Japanese models [1]. When a human draft is lightly polished, the underlying human structure survives and the model's signal weakens. When an LLM draft is heavily humanized, the same thing happens in reverse.

That symmetry is the core finding of this report: detection accuracy is a function of how much original human structure remains in the text, not of who or what wrote it.

Humanizer Performance: The 76.44% Evasion Result

Turnitin0's humanizer was tested on 174 essays generated by GPT-5.6-Sol, totalling 204,736 words. After humanization, the overall word-level evasion rate was 76.44% — meaning roughly three-quarters of previously flagged words stopped being flagged.

The discipline spread was wide. Education reached 100% evasion. English was lowest at 55.41%. That variance is worth stating plainly: humanization is not a uniform guarantee, and a student in English literature faces a harder detection environment than a student in education.

Turnitin0's humanizer preserves meaning, citations, headings, and.docx formatting, and it is built for text drafted with ChatGPT, Claude, or Gemini. It supports.docx and.txt files, English only, under 90 MB. Notably, 98.2% of humanizer orders are re-checked with Turnitin — the company's own workflow assumes you will verify the result rather than trust it.

How the Eight Tools Compare

The table below compares the eight tools using only vendor-stated claims and available user feedback. Strengths are listed honestly, including for competitors.

QuillBot appears in the material only as a secondary reference — described in a Reddit review as simple to use and free, but not the best for longer-form content. Its AI Checker is mentioned as one of several detectors in a Reddit tutorial.

What the Comparison Actually Shows

Two things stand out. First, most competitors publish performance claims with no independent verification. Humanize AI, Getsolved.ai, and humanizeai.pro all present strong numbers sourced entirely from their own homepages. Second, the tools that do publish honest limitations — turnitindetectorai.com explicitly states it "does not reproduce Turnitin's proprietary detection model" and "cannot predict an official Turnitin result" — are being straight with users about a real constraint.

Turnitin0 sits in a different category from the pure humanizers. Its checking service produces two downloadable PDFs per order: an AI detection report and a similarity/plagiarism report. Because it is non-repository, files are not added to Turnitin's student paper database, which means checking your own draft does not create a future similarity match against itself.

Real User Experience

Turnitin0's user base spans 20,000+ students across multiple countries. Reviews on the Trustpilot profile show a TrustScore of 4.3/5 with an "Excellent" label, 9 reviews, 89% of them 5-star, and no negative reviews.

Raini Dipré (CA) described the process as easy, fast, and efficient, with the report arriving much faster than expected. May Zin (SG) received a complete report in about 20 minutes and could download the AI and similarity reports together. Daniela Pellegrini (GB) has used Turnitin0 several times and found the humanizer helpful when revising. Shubham Pachauri (IN) found it quick and useful for checking and improving academic writing, and specifically liked the humanizer for making text sound more natural. Shawn Thakur (AU) called it easy to use and on time. "Encrypted" (GB) described it as the best site for Turnitin scans — authentic and simple. Taksh Patel (AU) called it 100% legit and working. A US reviewer known as "b c" used it for assignments, plagiarism checking, and general AI awareness.

One honest caveat: Trustpilot notes that the company has not recently invited customers to leave reviews, so the sample may not be representative, and all 9 reviews fall within the last 12 months. The 4.3/5 TrustScore is real but small.

Delivery Performance

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. For comparison, T-detector states 5–20 minutes and turnitindetectorai.com claims under 60 seconds for its independent detector — though the latter is not a Turnitin reproduction.

Practical Recommendations

Check before you submit, not after. The 47.54% accuracy figure on AI-polished human writing means the number your instructor sees may not reflect what you actually did. A pre-submission check gives you the real figure while you can still act on it.

Do not assume polishing is safe. If you used a model for grammar and flow on your own draft, you are in the highest-uncertainty category in this entire dataset.

Verify humanizer output. Turnitin0's own workflow re-checks 98.2% of humanizer orders. Follow that pattern.

Understand the confidence threshold. Turnitin0 displays any AI score below the 20% confidence threshold as *% rather than an exact percentage. This mirrors how Turnitin itself handles low-confidence results and is not a hidden score.

Use the non-repository option. Because Turnitin0 does not add files to Turnitin's student paper database, checking a draft does not create a self-match later.

Read the report the way your professor will. Turnitin flags at the sentence level, not just the document level, and AI writing highlights are not visible in the Similarity Report [1]. The two reports answer different questions.

Conclusion

The evidence points one way: Turnitin's AI detector is excellent at catching raw model output and excellent at clearing genuine human writing, but it is close to a coin flip on AI-polished human drafts, and it can be substantially evaded by humanization. That combination makes pre-submission verification the single highest-value step a student can take — and the Turnitin AI detector at Turnitin0 is the tool built specifically for it, with reports that match what your professor sees, a non-repository workflow that protects your future submissions, delivery under 15 minutes in 98% of cases, and a humanizer that preserves meaning, citations, headings, and formatting. Check your draft, read the report honestly, and submit with your eyes open.