Can AI detectors catch paraphrased or “humanized” AI text?
Short answer
Sometimes. It depends heavily on which detector and which rewriting tool. Paraphrasing collapsed early detectors: in a 2023 study, one detector's hit rate on paraphrased AI text fell from 70.3% to 4.6%. Newer detectors trained on rewritten text catch much more. In a June 2026 study, Pangram flagged 37 of 40 AI papers that had been rewritten to sound human, while GPTZero flagged 1.
No detector is immune. In an October 2026 preprint, a commercial humanizer (Undetectable AI) got more than 96% of AI rewrites past both Pangram 3.2 and GPTZero. A clean score on text that may have been humanized proves little, and humanizing work to hide AI use is still AI use.
What happens when AI text is paraphrased?
| Study | What they did | Result |
|---|---|---|
| Krishna et al., NeurIPS 2023 | Rewrote GPT-2 XL text with an 11-billion-parameter paraphraser (DIPPER) | DetectGPT's detection fell from 70.3% to 4.6% at a fixed 1% false positive rate. GPTZero, OpenAI's classifier and watermarking were also evaded. |
| Sadasivan et al., 2023 | Paraphrased watermarked AI text over and over | The watermark detector's hit rate fell from 99.3% to 9.7%, and human raters still judged most paraphrases good quality. |
| Weber-Wulff et al., 2023 | Ran AI text through QuillBot, then 14 detectors (2023 versions) | Only 26% correctly identified overall. |
| Liang et al., 2023 | Asked ChatGPT to rewrite its own essays "by employing literary language" | Detection of 31 college-application essays fell from up to 100% to up to 13% (seven 2023 detectors). |
| RAID, ACL 2024 | 11 kinds of attacks on 12 detectors | Effects varied by tool: swapping letters for look-alike characters cut Originality's accuracy from 85.0% to 9.3%, while GPTZero lost 0.3 points. One paraphrasing attack actually raised Originality's accuracy to 96.7%. |
| JAIT, 2026 | QuillBot paraphrasing of AI text, 9 detectors (small sample) | Turnitin's accuracy on Grok text fell to 45.7% and Grammarly's to 19.0%. |
An older defense still works in the lab: in the 2023 DIPPER study, searching a database of 15 million past AI outputs caught 80 to 97% of paraphrased text. But it needs the AI company to store and search everything its model writes, which no public detector can do.
What about AI humanizers?
Humanizers are tools sold specifically to make AI text pass detectors. Independent tests disagree sharply:
| Study | Humanizer | Share of humanized AI text each detector caught |
|---|---|---|
| Chicago Booth, Sep 2025 | StealthGPT | Pangram: 95% or more. OriginalityAI: about 95% of long texts, as low as 79% of short ones. GPTZero: missed around half or more (miss rates 25.5 to 77.3% by genre and model). |
| Vrije Universiteit Brussel, Jun 2026 | GPT-4o asked to make AI papers sound human (not a commercial humanizer) | Pangram 92.5%, Turnitin 50.0%, Copyleaks 22.5%, GPTZero 2.5%. |
| Notre Dame, Oct 2026 (preprint) | Undetectable AI v11 | Fewer than 4% stayed flagged by Pangram 3.2 and GPTZero. |
The Notre Dame authors tested Pangram 3.2. Pangram has since released version 4 (technical report dated July 2026), so results for the current model may differ.
What do the detector companies say?
- Turnitin added detection of AI paraphrasing tools on December 6, 2023, and announced "AI bypasser" (humanizer) detection on August 27, 2025, for English only. It doesn't name the tools it detects.
- Pangram says Pangram 4 classified humanized text as AI 97.67% of the time across 13 commercial humanizers (its own test, July 2026).
- GPTZero says it caught 93.5% of texts from 12+ paraphrasing tools on its own benchmark, versus 57.3% for Originality and 50.2% for Pangram v3 (January 2026).
- Originality.ai cites a benchmark that ranks GPTZero last instead of first. We couldn't find that benchmark published, which is one more reason to treat vendor comparisons on humanized text with caution.
Why do the results conflict so much?
- It's an arms race. Humanizers update to beat detectors; detectors retrain on humanized text. A result from six months ago may already be out of date.
- "Humanizer" covers very different tools, from a single chatbot prompt to dedicated paraphrasing models.
- Versions differ. Studies often test a detector version that has since been replaced.
- Vendors choose their own test sets and tend to win their own benchmarks.
What does this mean for teachers and students?
- For teachers: a low AI score doesn't prove a student wrote the text, because it may have been humanized. Process evidence, such as drafts, version history and a conversation about the work, matters more than any score. A fair checking process.
- For students: a humanizer doesn't change whether you used AI against the rules, and it can still be caught. Turnitin now looks for humanized text, and some detectors catch most of it. If AI was allowed, say how you used it. If it wasn't, the safe option is writing it yourself.
- For editors and clients: unusual phrasing, odd word swaps and awkward synonyms can be signs of a paraphrasing tool. Check the passages a detector highlights and ask the writer for drafts.
Questions people also ask
Can Turnitin detect QuillBot paraphrasing?
Turnitin added detection of AI paraphrasing tools in December 2023 and of AI "bypasser" (humanizer) tools in August 2025, for English only. Independent results are mixed: in a small 2026 study, Turnitin's accuracy on Grok text fell to 45.7% after QuillBot paraphrasing.
Do AI humanizers work?
Against some detectors, often. Against others, rarely. In 2025, a humanizer called StealthGPT got around half or more of AI passages past GPTZero, while Pangram still caught at least 95%. In 2026, Undetectable AI got more than 96% past both Pangram 3.2 and GPTZero. Results change with every detector and humanizer update.
Can AI Checker Plus detect humanized text?
It's built to flag AI text that was edited, paraphrased or run through a humanizer, and labels each passage AI-generated, AI-assisted or human-written. Like every detector, it can miss some rewritten text, so a human-looking result isn't proof that a person wrote it.
Sources
- Krishna et al., Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense (NeurIPS 2023)
- Sadasivan et al., Can AI-generated text be reliably detected? (arXiv, v4 Jan 2025)
- Weber-Wulff et al., Testing of detection tools for AI-generated text, IJEI (Dec 2023)
- Liang et al., GPT detectors are biased against non-native English writers (arXiv 2304.02819)
- Dugan et al., RAID benchmark, ACL 2024
- Makhmutova et al., Journal of Advances in Information Technology 17(3) (Mar 2026)
- Jabarian and Imas, Artificial Writing and Automated Detection, NBER Working Paper 34223 (Sep 2025)
- Van Vlasselaer et al., International Journal for Educational Integrity 22, article 16 (Jun 2026)
- Karr et al. (Notre Dame), AI detectors and AI-edited academic abstracts (arXiv, v3 Oct 2026)
- Turnitin: expands capabilities amid rising threats posed by AI bypassers (Aug 2025)
- Turnitin: AI writing detection model release notes
- Pangram 4 technical report (arXiv, Jul 2026)
- GPTZero: Detecting AI-humanized text (Jan 2026)