The Specification
This desk scores transparency, not performance. Bypass rates, evasion scores, and detector-beating percentages are excluded from scoring entirely, because the only public sources for them are vendor-run or anonymous and they contradict each other. What remains auditable is whether a vendor discloses who it is, publishes its method, and states its terms. The best AI humanizers of 2026 cannot be honestly ranked on public evidence, which is why bestaihumanizers.org catalogues them across vendor disclosure, methodology transparency, meaning-preservation research, pricing transparency, terms and acceptable use, and data handling.
Dimensions & weights
| Dimension | Weight | Measures / recomputation rule |
|---|---|---|
| Vendor disclosure | 25% | Mechanical: 60 points if the operator's identity is public, 40 if assessments carry a signed author. Anchored on the sampled record's finding that public assessments here are "vendor-run or anonymous". |
| Methodology transparency | 20% | Count, normalised: test set disclosed · detector list disclosed · version date disclosed · sample size disclosed. |
| Meaning-preservation research | 20% | Binary: is there a third-party (non-vendor) assessment of how faithfully the tool preserves the original meaning? Anchored on the Pangram Labs 19-tool study (Masrour, Emi and Spero, arXiv). |
| Pricing transparency | 15% | Binary: per-tier pricing and word quotas published on the vendor's own site. |
| Terms & acceptable use | 10% | Binary: a usage policy and disclaimers are published. |
| Data handling | 10% | Binary: the storage and use policy for submitted text is published. |
Scoring formula
Each dimension is normalised to 0-100, then combined. Any finding on this site is reproducible from these rules:
disclosure = 60·operator_public + 40·author_signed
method = minmax(count of 4 disclosed items) · 100
bypass_performance = excluded # conflicted/anonymous sources only
The exclusion rule (why no bypass scores)
Our October 2026 sampling of the public assessments, as tabulated by theairankings.com, found: HumanizerBench (October 2026) places Undetectable.ai 3rd of 14 and WriteHuman 1st of 14 — and per that site's own disclosure, WriteHuman operates it. AI Humanizer Benchmark (October 2026, operator unnamed) places the same two tools 11th of 11 (33.60, 1.8% bypass) and 7th of 11 (75.06, 97.6% bypass). Best AI Humanizer (Substack, unsigned) gives 79.7% and 74.1%. One tool is 3rd of 14 in a competitor-run source and last of 11 in an anonymous one. No figure from any of these sources appears anywhere on this site as fact; they appear only as attributed, conflicting claims — and none of them feeds the score.
What the peer-reviewed record measures
Two studies anchor this desk, cited with full attribution: Van Vlasselaer, Van Droogenbroeck and Spruyt (International Journal for Educational Integrity, June 2026) tested four detectors on 160 academic papers of at least 4,000 words and measured accuracy on humanized text at Pangram 92.5%, Turnitin 50.0%, Copyleaks 22.5%, GPTZero 2.5%. Jabarian and Imas (NBER working paper 34223, September 2025) found Pangram's error rate near zero across genres, lengths, and models. As theairankings.com notes, none of these studies ranks commercial humanizers against each other — each measures detectors against humanized text. That is precisely why detector-side data cannot be repurposed into a humanizer ranking here, however tempting the arithmetic.
Terminology
AI humanizer ≠ paraphrasing tool ≠ AI detector ≠ grammar checker. A humanizer rewrites text to evade AI detection. A paraphrasing tool rewrites for clarity without that goal (separate page). A detector estimates whether text was AI-generated — the opposite function, covered by related detector publications (see our note). A grammar checker corrects without restructuring. The senses are kept apart on every page of this site.
Ethics stance
This site does not provide, review, or endorse methods for evading detection. A humanizer changes how text reads; it does not add original thinking, check facts, or change who wrote the text — so passing AI-generated work through one does not make the work yours, and may violate academic integrity policies, employment originality clauses, and publisher submission rules. We catalogue the market and its data conflicts; the reader's obligations are the reader's to check.
Data coverage
The 2026 catalogue holds 4 tools — those present in the comparison inventory of theairankings.com (accessed October 2026). Tools named elsewhere in the market (Phrasly, StealthWriter, BypassGPT, Humanizer Pro, and others) are held out of the catalogue until their fields are verified; entries with zero independent verification are not added to pad the list. Per-tool transparency audits (operator identity, signed authors, published methods, pricing, terms, data handling) are in progress; until they complete, every catalogue entry shows index_score as pending and the 2026 edition is marked PROVISIONAL.
Disclosure
bestaihumanizers.org is part of a network of independent review publications. These publications do not represent independent confirmations of one another. The full list of related publications is in About.