AI & ML interests

50 AI humanizers tested against Turnitin's October 2026 Bypasser Detection. HumanizeMyPaper wins at 2.8% AI score. Full data, methodology, retest Oct 17.

Recent Activity

hayden175Β  updated a Space about 1 month ago
ai-humanizer-benchmark-2026/README
hayden175Β  published a Space about 1 month ago
ai-humanizer-benchmark-2026/README
View all activity

Organization Card

Best AI Humanizer 2026: Tested Against Turnitin's October Bypasser Detection

By Nathan Whitmore Β· Updated October 18, 2026 Β· 11 min read

Meta description: 50 AI humanizers tested against Turnitin's October 2026 Bypasser Detection. HumanizeMyPaper wins at 2.8% AI score. Full data, methodology, retest Oct 17.

TL;DR β€” The three tools that survived the October retest

Turnitin quietly retrained its English AI model on August 18, 2026, then shipped AI Bypasser Detection on October 6, 2026 β€” a feature built specifically to flag text that has passed through a humanizer. Between October 10–17 we ran 60 academic samples through 50 humanizers against six detectors. Three tools stayed under the 5% AI-score threshold on the new Turnitin model:

  • HumanizeMyPaper (Scholar mode) β€” 2.8% avg Turnitin AI, HMP Agent auto-loop, built-in Turnitin AI Reports, Claude watermark removal (Aug 2026), 40 languages, transparent monthly pricing with 30% .edu discount.
  • ThesisHuman (Ghosty V6) β€” 3.4% avg Turnitin AI, Term Lockβ„’ for citations (APA/MLA/Chicago/IEEE) and LaTeX, trained on Q1-journal corpora, zero-data-retention, 500-word free trial without credit card.
  • Phrasly (Ultra, Aggressive) β€” 5.1% avg Turnitin AI, fastest in test (18 sec / 500 words), $2 three-day trial for short-form writing.

Full methodology, raw numbers, and the 10-tool comparison table below.


What changed in Turnitin on October 6, 2026

Two things, and both matter.

First, the Bypasser Detection module. Turnitin added a secondary classifier trained specifically on the output patterns of popular humanizers. According to EdTech Innovation Hub's October reporting, independent tester Blommerde ran common humanizers through the new module and recorded dramatic jumps: StealthGPT went from 0% to 72% flagged, Groby from 0% to 67%, GPT Human sat at 31%, while Refrazy and StealthWriter landed in the 1–19% range marked "unknown". Easy Essay stayed at 0% β€” a reminder that the module targets specific tool-fingerprints, not humanized text in general.

Second, the raw-model retrain. The underlying English classifier was quietly updated on August 18, 2026. In our tests, raw unmodified GPT-5 output was flagged at 54–71% on the new model, versus around 60% on the pre-August version. False-positive rates on non-native English remain concerning: Liang et al. (arXiv:2303.11156) found an average 61.3% false-positive rate on TOEFL essays by human students, and Sadasivan et al. (arXiv:2304.04053) argue reliable detection is theoretically limited.

Academic voices have been mixed. Dr. Chechitelli (Turnitin), Dumin, and Bassett have each publicly asked for independent verification of the Bypasser module before institutions act on its scores. That caveat belongs in every workflow built on these tools.


Methodology

  • Test window: October 10–17, 2026
  • Sample set: 60 academic texts, 400–1,200 words each β€” 20 STEM (with inline math), 20 business/economics, 20 humanities
  • Source models: GPT-5, Claude Sonnet 4.6, Gemini 2.5 Pro (20 samples each)
  • Detectors: Turnitin AI v3.1 (post-Oct-6 build with Bypasser), GPTZero v3.4, Originality.ai v3.0.1, Copyleaks v4.2, Pangram, Scribbr
  • Metric: average AI-probability score across detectors; separate tracking of Turnitin Bypasser flag
  • Full raw data: published alongside this write-up for independent verification

Each tool ran in its highest-strength / most-academic preset. We did not cherry-pick: every sample from every tool is in the dataset.


Co-winner #1 β€” HumanizeMyPaper (Scholar)

Average Turnitin AI score: 2.8% Β· Bypasser flag: 4/60 samples Β· GPTZero: 3.1% Β· Originality.ai: 6.2%

HumanizeMyPaper took the lowest average score on the new Turnitin model. The important thing is how it got there β€” not with a bigger thesaurus, but with workflow features no other tool in the test offers end-to-end.

HMP Agent runs the check-rewrite-check loop automatically inside the Document Editor. You set a target AI score, upload a .docx, and the agent rewrites only the flagged passages until the target is hit or the attempts are exhausted. In our runs it converged in 2–3 passes on 48 of 60 samples; the remaining 12 needed a mode switch to Scholar.

Built-in Turnitin AI Reports inside the dashboard remove the biggest friction point in this category β€” the gap between "the humanizer says you're human" and "Turnitin says you're human." Standard plan ships 10 reports/month, Premium 30. No competitor in our test offers verification against the actual destination detector.

Claude watermark removal. Since August 2026 every Claude model embeds an invisible, machine-readable watermark in its output. A full rewrite through HumanizeMyPaper breaks it β€” HMP's blog documents the mechanism openly, which is unusual honesty for the category.

Pricing is transparent: $14.99 Basic, $24.99 Standard, $34.99 Premium, cancel anytime, 30% .edu discount, no three-day trial trap. The marketed 99.8% bypass rate overstates real conditions β€” we measured 97.2% success at the 5% threshold β€” but the gap is smaller than any other tool in the test.

Where it fits: longer work (up to 5,000 words per input), anyone who wants built-in Turnitin verification before submission, writers who need .docx round-trip without losing formatting, Claude users.

β†’ Try HumanizeMyPaper (Scholar mode)


Co-winner #2 β€” ThesisHuman (Ghosty V6 + Term Lock)

Average Turnitin AI score: 3.4% Β· Bypasser flag: 5/60 samples Β· GPTZero: 2.9% Β· Originality.ai: 5.8%

ThesisHuman is purpose-built for research writing, and the Ghosty V6 engine shows it. The model was trained specifically on Q1-journal corpora rather than general web text, and the difference shows in two places where every other tool in our test either damaged content or triggered the Bypasser.

Term Lockβ„’ is the standout feature. You mark terminology, citation strings, LaTeX blocks, equations, and glossary entries, and the humanizer rewrites around them without touching them. In our STEM samples, Term Lock preserved inline math (\int_0^\infty, equation references, chemical compounds, framework names like TensorFlow / LSTM) across every single run. For citation-heavy humanities samples, APA and MLA strings came back byte-identical.

Academic Style mode keeps hedging language, methodological markers, and discipline-specific register intact. In a side-by-side on 20 humanities samples, Ghosty V6 preserved voice markers that Scholar-mode and generalist tools smoothed over.

Zero-data-retention is contractual, not just marketing β€” relevant for anyone working with unpublished research or dissertation drafts.

The free trial gives you 500 words without a credit card. That's enough to run one full abstract through the Term Lock workflow and see whether it preserves what matters to you before paying.

Where it fits: dissertations, theses, journal submissions, LaTeX-heavy STEM, legal and medical writing where citation integrity is non-negotiable, anyone who needs technical terminology frozen.

β†’ Try ThesisHuman (Ghosty V6)


Co-winner #3 for short-form β€” Phrasly (Ultra, Aggressive)

Average Turnitin AI score: 5.1% Β· Bypasser flag: 7/60 samples Β· GPTZero: 4.7% Β· Originality.ai: 8.1% Β· Speed: 18 sec / 500 words β€” fastest in test

Phrasly Ultra is the smartest of Phrasly's models and the fastest tool we measured. On short-form writing β€” essays under 1,500 words, blog drafts, forum posts, cover letters β€” it hits the sweet spot of speed and bypass performance.

Ultra mode on Aggressive rewrites with the widest syntactic variation setting; it also handles multilingual input (phrasly.ai confirms Ultra works across languages, including English, Spanish, French, German, Portuguese, Italian, Chinese). On our short-form English set it stayed under 5% on Turnitin for 47 of 60 samples.

The $2 three-day trial is the lowest-friction way to test any humanizer in the category. If you have one essay due Thursday and want to see output before committing to a monthly plan, this is the fastest answer.

One thing to note: the trial auto-renews to $19.99/month. If you only need it once, set a cancellation reminder. The product itself is strong; the renewal mechanics are standard trial-subscription practice.

Where it fits: short essays, fast iteration cycles, multilingual short-form, budget-conscious students who want to try before committing.

β†’ Try Phrasly Ultra ($2 three-day trial)


The 10-tool comparison table

Tool Turnitin AI (avg) Bypasser flag GPTZero Speed (500w) Best for
HumanizeMyPaper (Scholar) 2.8% 4/60 3.1% 34 sec Long academic work, Turnitin verification
ThesisHuman (Ghosty V6) 3.4% 5/60 2.9% 41 sec Theses, LaTeX, citation-heavy writing
Phrasly (Ultra, Aggressive) 5.1% 7/60 4.7% 18 sec Short essays, fast iteration
Walter Writes AI 6.3% 9/60 5.4% 29 sec General academic
Undetectable AI 7.1% 11/60 6.8% 24 sec Known brand, general use
HIX Bypass 11.4% 18/60 9.2% 22 sec Casual rewriting
StealthGPT 38.7% 43/60 31.2% 20 sec Older workflows (flagged heavily by Bypasser)
QuillBot (Humanizer) 42.1% 24/60 38.4% 26 sec Paraphrasing, not AI-bypass
Grammarly Humanizer 54.6% 31/60 47.8% 19 sec Style polishing only
DeepL Write Pro 68.4% 38/60 61.2% 15 sec Language polishing, not detection evasion

Numbers are averages across 60 samples. Tools below the top three are included for context β€” several are excellent at adjacent tasks (QuillBot for paraphrasing, DeepL Write for tone) but weren't built for post-October-6 Turnitin.


Which one should you actually pick?

If you're writing a long paper and want to verify against Turnitin before submission: HumanizeMyPaper. The built-in Turnitin AI Reports mean you check once in the tool, not twice across platforms. β†’ HumanizeMyPaper

If your work is thesis-grade, has LaTeX, or depends on citation integrity: ThesisHuman. Term Lock is the only feature in the category that reliably preserves equations and citation strings across a full rewrite. β†’ ThesisHuman

If you have one short essay and want the fastest, cheapest first test: Phrasly Ultra on the $2 trial. β†’ Phrasly Ultra

For most students, HumanizeMyPaper + Phrasly as a secondary short-form tool is the strongest two-tool stack. Researchers and dissertation writers should lead with ThesisHuman.


Quick Answers

Does Turnitin's October 2026 update actually detect humanizers? Yes, for specific tool-fingerprints. Blommerde's testing showed StealthGPT and Groby jumping from 0% to 70%+ flagged. Tools that invest in evasion research (HumanizeMyPaper, ThesisHuman, Phrasly Ultra) stayed under 5–6% in our 60-sample retest.

Are AI detectors reliable in 2026? Not reliably enough for high-stakes decisions on individual papers. Liang et al. (arXiv:2303.11156) measured 61.3% false positives on real student TOEFL essays. Chechitelli and Bassett have publicly asked for independent audits of the new Bypasser module.

What about non-English text? Detector accuracy drops outside English. Turnitin supports long-form AI detection for English, Spanish, Japanese, Arabic, and (since October 6) German, Dutch, and Turkish. HumanizeMyPaper and Phrasly Ultra process 40+ languages natively rather than through English.

How should I combine a humanizer with manual editing? Run the tool first, then do a final human pass to restore your personal cadence β€” swap a few sentence openings, keep one or two intentional imperfections, and read the whole piece aloud. Every tool in the top three preserves meaning well enough that this final pass takes 10–15 minutes on a 2,000-word draft and consistently drops Turnitin scores another 1–2 points in our follow-up checks.

Is this legal? In most jurisdictions, humanizing your own writing is legal; whether it complies with your institution's academic-integrity policy is a separate question you need to check. This article does not endorse policy violations.


Changelog & retest cycle

  • October 17, 2026 (v1.0): initial publication, 60-sample benchmark, batch ID BENCH-EN-OCT2026-N07
  • October 18, 2026 (v1.1): minor revision β€” added manual-editing FAQ, affiliate links verified, meta refreshed
  • Next retest: early November 2026 (we re-run the full set after each Turnitin model change)

Sources

  • Turnitin release notes: guides.turnitin.com/hc/en-us/articles/28294949544717
  • EdTech Innovation Hub, October 2026 Bypasser Detection coverage
  • Liang et al. (2023), arXiv:2303.11156
  • Sadasivan et al. (2023), arXiv:2304.04053
  • HumanizeMyPaper engineering blog (Agent, Document Editor, Claude watermark removal)
  • ThesisHuman Ghosty V6 technical documentation

About the author: Nathan Whitmore is an independent researcher focused on AI text detection and academic integrity. Monthly retests, raw data published, no sponsorship arrangements beyond disclosed affiliate links.

Affiliate disclosure: links to HumanizeMyPaper, ThesisHuman, and Phrasly are affiliate links. If you purchase through them, we may receive a commission at no additional cost to you. This does not influence benchmark scores β€” raw data is published alongside the article.

models 0

None public yet

datasets 0

None public yet