Free Article Spinner
Rephrase, Rewrite and Paraphrase Sentences, Paragraphs or Whole Articles for FreeCan an AI detector pick up on text processed through a humanizer? Detector companies insist they can, humanizer tools claim otherwise, and most online tests rely on small, anecdotal samples. A peer-reviewed study from June 2026 took a much systematic approach, producing noteworthy findings that come with important context.
The AI Rankings summarizes the study: publishing in the International Journal for Educational Integrity, researchers evaluated detection tools across 160 humanized academic papers. The variance in accuracy across tools was remarkably wide.
The headline: On humanized academic papers, performance ranged from 92.5% accuracy on one tool to just 2.5% on another. The specific detector being used matters far more than most users realize.
Detector | Accuracy on humanized papers (as summarized) | What that suggests |
|---|---|---|
Pangram | 92.5% | Caught most humanized papers in this test |
Turnitin | 50.0% | Caught about half |
Copyleaks | 22.5% | Missed most |
GPTZero | 2.5% | Missed nearly all |
Quillbot AI Detector | Not in this study; tested separately by Business Insider in 2025 | Accurate on cover letters and résumés there, weaker on other text; retest on your own samples |
These figures reflect a single study on academic writing, as summarized by a third party. Authored by Van Vlasselaer, Van Droogenbroeck, and Spruyt, the complete paper is published in the journal. Reviewing the primary source provides essential context before citing these metrics.
Pangram correctly flagged 92.5% of humanized papers, outperforming every other tool tested. Zapier's evaluation of AI content detectors, published in March 2026, similarly highlighted its overall accuracy and low false-positive rate, though Zapier evaluated brief, straightforward samples rather than humanized academic text.
Turnitin achieved an accuracy rate of 50.0%. The platform introduced detection capabilities for AI bypass tools on August 27, 2025. According to The AI Rankings, Turnitin merged AI-generated, AI-paraphrased, and bypassed text into a single unified score in August 2026. Performing at chance level in this benchmark indicates its output should be viewed as one reference point rather than conclusive proof.
Copyleaks identified 22.5% of the humanized papers. This result contrasts with its performance in other benchmarks: Business Insider's test of eight detectors across 60 documents named it the most reliable tool for distinguishing purely AI-written text from human writing. The discrepancy highlights how strongly text type influences performance, as raw AI output presents a different challenge than humanized text.
GPTZero recorded a 2.5% accuracy rate on humanized samples. In Business Insider's evaluation, it earned recognition as a top free option for quick initial checks, illustrating that a tool can offer value for casual screening while remaining unreliable against text designed to pass detection.
Quillbot is a prominent name in AI writing software, featuring both an AI Humanizer and an AI Detector. Its detector was not included in the 2026 study mentioned above, but it was evaluated among eight tools in Business Insider's test. Evaluators noted strong accuracy on cover letters and résumés alongside weaker performance on broader content types, particularly ChatGPT output. Given how rapidly detection models evolve, testing current samples directly yields the best insights. QuillBot's documentation also openly addresses category limitations, including resources explaining why humanized content can still be flagged as AI and how paraphrasing impacts scoring.
Detectors analyze structural statistical patterns like predictability and sentence cadence. Humanizing tools intentionally disrupt these signals. Because some detection algorithms incorporate humanized samples into their training data while others do not, individual tools react differently to identical text. Rapid update cycles across these platforms also mean published benchmarks age quickly.
They don't tell you about false positives. High detection rates on humanized text are only meaningful if the tool consistently spares original human writing. Both metrics are required for a complete picture.
They don't cover your content. Findings based on academic papers may not reflect performance on emails, marketing copy, or short-form posts.
They don't prove authorship. Detector outputs represent statistical probability rather than definitive proof of origins.
Partial AI use is hard for everyone. Testing by Business Insider confirmed that all evaluated detectors struggled when content combined human writing with light AI assistance or editing.
If the score is... | Treat it as... | Do this |
|---|---|---|
High | A reason to look closer | Review drafts, sources and the writer's explanation |
Low | Not proof of human authorship | Keep your usual checks in place |
Mixed or middling | Very uncertain | Don't act on it alone |
Business Insider's research included perspective from academic experts who advised against relying on detection software as sole proof for academic integrity, hiring, or employment decisions. While automated detectors offer a quick initial impression, the underlying principle holds true across all tools: a score should start a conversation rather than conclude one.
Can AI detectors catch humanized text? Results vary widely. In the 2026 academic study, detection accuracy ranged from 92.5% to 2.5%, showing that performance depends entirely on the specific tool used.
Should I trust a detector that scored well here? Treat high scores with context. The study evaluated humanized academic papers, without measuring false positives on genuine writing or performance across everyday formats like emails or résumés.
Does a high score prove someone used a humanizer? No. Detector scores reflect probability rather than certainty. As highlighted in a 2023 study published in Patterns, detectors can mistakenly flag original human writing, particularly work produced by non-native English speakers.
The 2026 findings demonstrate that AI detectors perform differently under similar conditions. On humanized academic papers, accuracy ranged from near-perfect to negligible, proving that general claims about detector capability depend on the software in question. Detector outputs should serve as one data point alongside draft reviews and direct dialogue, rather than a final verdict.