Which AI detectors detect AI-paraphrased text? The 16 vendors that claim it, the 5 that report it as a label or setting of its own, and what 11 studies measured
Sixteen of the 72 AI detectors in the index state on their own pages that they detect AI text after paraphrasing, humanizing or AI editing, by the humanized text statement field of every card read 8 and 9 November 2026 in the data file, and five of the sixteen report it as something other than a higher AI score: TruthScan's API returns a Paraphrase label beside Human and AI (truthscan.com, read 9 November 2026), QuillBot's results separate "Generated and refined by AI" from "Written by a human and refined by AI" (quillbot.com, read 8 November 2026), Pangram highlights AI-edited segments (pangram.com, read 8 November 2026), Copyleaks sells an "Extra Sensitive" model for humanizer output (copyleaks.com, read 8 November 2026), and Turnitin's release notes of 6 December 2023 and 27 August 2025 add detection of AI paraphraser and bypasser tools, for English only (guides.turnitin.com, read 8 November 2026). Eleven of the 35 study records scored a detector on paraphrased or humanized AI text; the oldest, Weber-Wulff 2023, recorded 15% average accuracy across 14 tools on 9 ChatGPT texts rewritten by QuillBot, against 63% and 70% on the two unedited ChatGPT classes (DOI 10.1007/s40979-023-00146-z, read 9 November 2026).
"Detects paraphrased AI" can mean three things, and the first table records which one each vendor states: a robustness claim reported inside the ordinary AI score, a separate label or highlight that names the paraphrased or AI-edited state, or a setting the user chooses for humanizer output. Every row quotes the card's statement with the vendor page and read date; the second table covers the cards whose statement makes no such claim or says the opposite. This page is one of the questions answered from the index; the adversarial robustness ranking orders the cards by the measured drop under paraphrase, and the AI detector register carries every card.
| Detector | Vendor statement, as recorded on the card | Reported as | Source | Read date |
|---|---|---|---|---|
| TruthScan | "Detect AI-generated content even after it's been paraphrased, rewritten, or humanized"; the API returns a Paraphrase label beside Human and AI | Separate label | truthscan.com | 9 November 2026 |
| QuillBot | Results distinguish "Generated by AI", "Generated and refined by AI", "Written by a human and refined by AI" and "Written by a human"; no claim about humanizer output | Separate categories | quillbot.com | 8 November 2026 |
| Pangram | "Pangram can detect AI-generated text even after it has been 'humanized'"; "AI Assistance: Pangram will highlight segments that have been edited by AI rather than fully AI-generated" | Separate highlight | pangram.com | 8 November 2026 |
| Copyleaks | "Paraphrased AI Detection"; "Extra Sensitive: Our most sensitive model was designed to flag AI text that was put through a 'humanizer' or text spinner" | Setting (sensitivity model) | copyleaks.com | 8 November 2026 |
| Turnitin | "Identify instances where AI-generated text may have been modified by AI paraphraser or bypasser (also called humanizers) tools"; paraphrase detection added 6 December 2023, bypasser detection 27 August 2025; English only | Separate indication in the report | guides.turnitin.com | 8 November 2026 |
| GPTZero | GPTZero 4o "deliberately tackles ... AI-generated text that has been paraphrased or mixed with human writing"; the FAQ still says the classifier "is not trained to identify AI-generated text after it has been heavily modified" | Within the AI score | gptzero.me | 8 November 2026 |
| Originality.ai | "Trained on Adversarial Data"; HumanizerBench retests 89.52% (5.0) and 92.38% (5.0 Max) on 420 humanized AI outputs, with no human controls | Within the AI score | originality.ai | 8 November 2026 |
| Winston AI | "Identify AI-generated writing even when its structure and wording have been substantially modified"; pricing row "Paraphrased Content Detection" | Within the AI score | gowinston.ai | 8 November 2026 |
| Sapling | "Significantly higher detection of humanized AI text", improved by 3.9% absolute (changelog 17 September 2026); "edited AI writing can be missed" | Within the AI score | sapling.ai | 8 November 2026 |
| JustDone | "Patterns that survive paraphrasers and grammar tools, so 'humanized' AI text is still caught"; blog: flags 60% of AI-edited human drafts | Within the AI score | justdone.com | 8 November 2026 |
| Proofademic | "Paraphrase Shield looks for AI-like patterns that can remain after text is processed through paraphrasing or humanizer tools"; "an active area of work, not a solved problem" | Within the AI score | proofademic.ai | 8 November 2026 |
| Undetectable AI | "Designed to identify rewritten or obfuscated AI content, including material modified by paraphrasing tools"; the same vendor sells a humanizer | Within the AI score | undetectable.ai | 9 November 2026 |
| Surfer | "Flag content that is likely to be AI-generated, including content altered by writing tools or AI rewriters"; the same page offers to rewrite AI content to pass detectors | Within the AI score | surfer.ai | 9 November 2026 |
| Compilatio | "Identify passages that may have been written or reworded by a generative AI tool"; "Altered text detection" for manipulations that conceal AI content | Within the AI score | compilatio.net | 9 November 2026 |
| Quetext | "Yes, Quetext's AI Detector ... flag[s] content that may have been paraphrased or lightly edited from an AI draft" | Within the AI score | quetext.com | 9 November 2026 |
| Codeleaks | "The ability to detect AI text paraphrased or interspersed with human text", footnoted to the text detector's English datasets; no statement about modified code | Within the AI score | copyleaks.com | 9 November 2026 |
Statements as the humanized text statement field of each card records them, read on the date shown; quotation marks mark the vendor's words.
Which detectors say they catch paraphrased or humanized AI text?
Sixteen, and the five that report it as a label, highlight, indication or setting are the ones a reader can check in the output. TruthScan's documentation describes a result label of Human, AI or Paraphrase and a score from 1 to 100 (truthscan.com, read 9 November 2026). Pangram's AI Assistance highlight marks AI-edited segments, and Epoch AI 2026 records that both Pangram and GPTZero return a "Mixed" verdict, which that study counted as a miss on AI text and would have counted as a false positive on human text, where neither returned one on 495 human passages (epoch.ai, read 9 November 2026). Turnitin limits its paraphraser and bypasser detection to English submissions (guides.turnitin.com, read 8 November 2026), and Copyleaks' Extra Sensitive model is a user setting rather than a label (copyleaks.com, read 8 November 2026).
The other eleven state robustness inside the ordinary score, and four claimants sell the opposite service on the same site: Undetectable AI, Surfer, Quetext and TruthScan's owner, Undetectable Inc. (undetectable.ai; surfer.ai; quetext.com; truthscan.com, read 9 November 2026).
What did studies measure on paraphrased AI text?
Eleven records, and the drops are large wherever a paraphraser was applied once. Weber-Wulff 2023 recorded 15% average binary accuracy across 14 tools on the QuillBot-paraphrased class of 9 documents, with 71.4% of paraphrased documents undetected; Compilatio missed 7 of 9 and Writer and DetectGPT each classified 1 of 9 correctly (DOI 10.1007/s40979-023-00146-z, read 9 November 2026). Orenstrakh 2023 ran 10 QuillBot-paraphrased ChatGPT submissions: Copyleaks fell from 100% to 50%, Originality.ai from 100% to 40%, the GPT-2 Output Detector from 100% to 60%, GPTZero from 70% to 20%, and GLTR held at 100% (arXiv 2307.07411, read 9 November 2026). Krishna 2023 found one DIPPER paraphrase cut GPTZero's detection of GPT-2 XL text from 13.9% to 1.2% and DetectGPT's from 70.3% to 4.6% at a fixed 1% false positive rate (arXiv 2303.13408, read 9 November 2026); Sadasivan 2023 recorded DetectGPT's AUROC falling from 96.5% to 59.8% after one T5 round on 1,000 passages (arXiv 2303.11156, read 9 November 2026).
The newer records separate the vendors. DAMAGE 2025 recorded GPTZero's true positive rate at a fixed 5% false positive rate falling from 99.73% to 60.04% after humanizing, while Pangram Labs' own DAMAGE model held 98.26% and its LLM Baseline 95.74% (ACL Anthology 2025.genaidetect-1.9, read 9 November 2026). Jabarian and Imas 2025 rewrote 1,992 AI passages per model once with StealthGPT: GPTZero's false negative rate rose to between 0.2552 and 0.7730, Originality.ai's to between 0.0227 and 0.2073, and Pangram's stayed between 0.0000 and 0.0500 (NBER w34223, read 9 November 2026). Van Vlasselaer 2026 placed 37 of 40 humanised papers in range for Pangram, against 50.0% accuracy for Turnitin, 22.5% for Copyleaks and 2.5% for GPTZero (DOI 10.1007/s40979-026-00226-w, read 9 November 2026). Alshammari and Rao 2025 recorded average AI scores on 49 humanized DeepThink paraphrases of 71% for Copyleaks, 58% for QuillBot and 52% for GPTZero (arXiv 2507.17944, read 9 November 2026); Kar 2024 read one QuillBot-paraphrased ChatGPT article as 0 AI on Copyleaks, 29 on GPTZero, 32.31 on ZeroGPT, 90 on QuillBot and 100 on Sapling and Undetectable AI (DOI 10.1177/02537176241247934, read 9 November 2026). RAID 2024 is the exception: under its paraphrase attack Originality.ai's accuracy at 5% false positives rose from 85.0 to 96.7 while GPTZero's fell from 66.5 to 64.0 (DOI 10.18653/v1/2024.acl-long.674, read 9 November 2026). Perkins 2024 folded paraphrase into six adversarial techniques (DOI 10.1186/s41239-024-00487-w, read 9 November 2026); the false negative rate ranking carries the miss rates these records produce.
Which detectors make no such claim, or say the opposite?
Twenty-six cards. Three text vendors state the limit rather than the claim, and the question on a 100 percent accurate detector collects such statements: Grammarly says "lightly edited AI-generated text may evade detection" (grammarly.com, read 8 November 2026), Scribbr says its detector "can make mistakes if the AI output was ... edited or paraphrased" (scribbr.com, read 8 November 2026), and ZeroGPT's terms state that it "does not guarantee that text processed by a humanization, paraphrasing, rewriting, or similar tool will be classified as human-written" (zerogpt.com, read 8 November 2026). Phrasly distinguishes AI-generated text from human writing that used AI tools and makes no claim about humanizer output while selling its own humanizer (phrasly.ai, read 8 November 2026). Six cards carry no statement in the field and are left out of the table: Walter Writes, Hive, WasItAI, Reality Defender, Deepware and Illuminarty, the last five of them image, video or voice detectors (walterwrites.ai, read 8 November 2026; the other five cards link their product pages).
Cards whose statement makes no paraphrase claim (20 of the 26)
| Detector | Statement on paraphrased or edited text, as recorded | Source | Read date |
|---|---|---|---|
| Grammarly | "Lightly edited AI-generated text may evade detection" | grammarly.com | 8 November 2026 |
| Scribbr | "Can make mistakes if the AI output was prompted to be less predictable or was edited or paraphrased" | scribbr.com | 8 November 2026 |
| Phrasly | Distinguishes fully AI text from human writing that used AI tools; no claim to detect humanizer output; sells a humanizer | phrasly.ai | 8 November 2026 |
| ZeroGPT | "Paraphrasing can reduce some signals"; terms: no guarantee that humanized or paraphrased text is classified as human by ZeroGPT or another detector; sells a humanizer | zerogpt.com | 8 November 2026 |
| GLTR | Research note: adversaries might change sampling parameters per word or sentence; "an adversarial sampling scheme would lead to worse text" | gltr.io | 9 November 2026 |
| AIDetectPlus | "Heavily edited AI text" named as one of two ways results go wrong; sells a humanizer | aidetectplus.com | 9 November 2026 |
| GPT-2 Output Detector | 2019 README: "adversaries can evade detection via finetuning from released models"; nucleus sampling outputs hardest to classify | github.com | 9 November 2026 |
| Codequiry | Code: "students who heavily modify AI-generated code may reduce detection confidence, but distinctive AI patterns often remain" | codequiry.com | 9 November 2026 |
| Writer | Archived page prompted users to "edit your text until there's less detectable AI content"; withdrawn 22 December 2025 | web.archive.org | 9 November 2026 |
| AI or Not | "Models are updated frequently for new adversarial attacks" (image page); no statement on paraphrased text | aiornot.com | 9 November 2026 |
| Grammarly Authorship | Tracks typing and pasting; categorises pasted text by source; no statement on paraphrased text | support.grammarly.com | 9 November 2026 |
| Turnitin Clarity | Flags measure writing time, revision effort and pasted text; AI results only with the Originality add-on; no statement on paraphrased text | guides.turnitin.com | 9 November 2026 |
| SynthID | Watermark "robust to some transformations", including "mild paraphrasing"; confidence "greatly reduced" when text is thoroughly rewritten or translated | ai.google.dev | 9 November 2026 |
| Smodin | No claim; the same page sells "Humanize AI text so it reads naturally, then re-check it here" | smodin.io | 9 November 2026 |
| BrandWell | Archived page: finds AI "even if the content has been mixed with human writing"; no statement on paraphrased text; sold WriteWell to pass detection | web.archive.org | 9 November 2026 |
| DetectGPT | Paper: AUROC above 0.8 with a quarter of the text replaced; cites Sadasivan 2023 and Krishna 2023 on paraphrase evasion as future work | arxiv.org | 9 November 2026 |
| Sightengine | Images: robust to re-encoding, compression and resizing; no text analysis | sightengine.com | 9 November 2026 |
| Sensity | "Robust adaptability against evolving threats"; no text detection | sensity.ai | 9 November 2026 |
| Pindrop Pulse | "Over 90% detection for new or unseen deepfake generation engines"; no text detection | pindrop.com | 9 November 2026 |
| Resemble Detect | Attack coverage table for voice clone, replay, face swap and partial edit; no text detection | resemble.ai | 9 November 2026 |
Statements as the humanized text statement field of each card records them, read on the date shown.
No vendor among the sixteen publishes a false positive rate for its paraphrase or humanizer detection on human-written text, and the one vendor benchmark, Originality.ai's 420 humanized outputs, ran with no human controls.
Questions about detecting paraphrased AI text
Can AI detectors detect paraphrased AI text?
16 of 72 vendors claim it and 11 of 35 study records measured it, read 9 November 2026; a single paraphrase cut GPTZero's detection from 13.9% to 1.2% in Krishna 2023 and from 99.73% to 60.04% after humanizing in DAMAGE 2025, while Pangram held 98.26% in DAMAGE 2025 and 37 of 40 humanised papers in Van Vlasselaer 2026.
Which AI detectors label paraphrased text separately?
5 of the 16: TruthScan returns a Paraphrase label, QuillBot separates "Generated and refined by AI" from "Written by a human and refined by AI", Pangram highlights AI-edited segments, Turnitin indicates AI paraphraser and bypasser use in English submissions, and Copyleaks offers an Extra Sensitive model, by their pages read 8 and 9 November 2026.
Does Turnitin detect QuillBot paraphrasing?
6 December 2023 is the date Turnitin's release notes give for paraphrase detection and 27 August 2025 for bypasser detection, English only, with no tools named (guides.turnitin.com, read 8 November 2026). Before that date, Weber-Wulff 2023 found all 10 of Turnitin's 36 missed AI documents among the manually edited and QuillBot-paraphrased texts (DOI 10.1007/s40979-023-00146-z).
Which detector held up best against paraphrasing in a study?
3 records tested Pangram on paraphrased or humanised text and it held up best in each: false negative rate at most 0.0500 after StealthGPT rewrites in Jabarian and Imas 2025, 98.26% true positives after humanizing in DAMAGE 2025 (the vendor's own research model), and 37 of 40 humanised papers in range in Van Vlasselaer 2026, read 9 November 2026. GLTR held 100% on 10 QuillBot-paraphrased texts in Orenstrakh 2023.
Which detector vendors also sell a humanizer?
4 of the 16 claimants' statement fields record a humanizer or paraphraser sold or linked on the same site, read 9 November 2026: Undetectable AI, Surfer, Quetext and TruthScan (through its owner, Undetectable Inc.). Among the non-claimants, Phrasly, ZeroGPT, Smodin, AIDetectPlus and BrandWell's archived page do the same.
Sources
All read 9 November 2026 unless an earlier date is shown; the source of every row is linked in the tables.
- TruthScan AI detector: https://truthscan.com/ai-detector
- QuillBot AI detector blog: https://quillbot.com/blog/quillbot-tools/ai-detector/ (read 8 November 2026)
- Pangram: https://www.pangram.com/ (read 8 November 2026)
- Copyleaks testing methodology: https://copyleaks.com/ai-detector/testing-methodology (read 8 November 2026)
- Turnitin AI writing detection model release notes: https://guides.turnitin.com/hc/en-us/articles/28294949544717-AI-writing-detection-model (read 8 November 2026)
- GPTZero 4o launch post: https://gptzero.me/news/introducing-gptzero-4o/ (read 8 November 2026)
- Originality.ai 5.0 launch post: https://originality.ai/blog/introducing-originality-5-0-and-max (read 8 November 2026)
- Winston AI detector: https://gowinston.ai/ai-content-detector/ (read 8 November 2026)
- Sapling detector API documentation: https://sapling.ai/docs/api/detector/ (read 8 November 2026)
- JustDone AI detector: https://justdone.com/ai-detector (read 8 November 2026)
- Proofademic Paraphrase Shield: https://proofademic.ai/paraphrase-shield/ (read 8 November 2026)
- Undetectable AI detector: https://undetectable.ai/ai-detector
- Surfer AI detector: https://surfer.ai/ai-content-detector/
- Compilatio API: https://www.compilatio.net/en/plagiarism-detection-api/
- Quetext AI detector: https://www.quetext.com/ai-content-detector
- Copyleaks Codeleaks: https://copyleaks.com/codeleaks
- Grammarly AI detector: https://www.grammarly.com/ai-detector (read 8 November 2026)
- Scribbr AI detector: https://www.scribbr.com/ai-detector/ (read 8 November 2026)
- ZeroGPT FAQ: https://www.zerogpt.com/faq (read 8 November 2026)
- Phrasly AI detector: https://phrasly.ai/ai-detector (read 8 November 2026)
- SynthID safeguards documentation: https://ai.google.dev/responsible/docs/safeguards/synthid
- Epoch AI 2026 data insight: https://epoch.ai/data-insights/ai-detectors-false-negatives
- Weber-Wulff 2023: https://doi.org/10.1007/s40979-023-00146-z
- Orenstrakh 2023: https://arxiv.org/abs/2307.07411
- Krishna 2023: https://arxiv.org/abs/2303.13408
- Sadasivan 2023: https://arxiv.org/abs/2303.11156
- DAMAGE 2025: https://aclanthology.org/2025.genaidetect-1.9/
- Jabarian and Imas 2025: https://www.nber.org/papers/w34223
- Van Vlasselaer 2026: https://doi.org/10.1007/s40979-026-00226-w
- Alshammari and Rao 2025: https://arxiv.org/abs/2507.17944
- Kar 2024: https://doi.org/10.1177/02537176241247934
- RAID 2024: https://doi.org/10.18653/v1/2024.acl-long.674
- Perkins 2024: https://doi.org/10.1186/s41239-024-00487-w
Revision log: 2026-11-09: page created from the humanized text statement field of 42 cards and 11 study records at index version 2026.11.7.
Related pages: Which AI detectors were tested on GPT-4 text? The 13 cards scored on GPT-4 family output in 7 study records, with every result and its base, Which AI detectors have the most false positive studies? GPTZero with 11 records, the GPT-2 Output Detector with 8, and every human-text figure the index holds for 15 text cards, AI detector false negative rates, ranked and Questions about AI detectors, answered from the index.