Benchmark guide · September 2026
AI Humanizer Benchmarks Compared: Why Different Tests Pick Different Winners
AI humanizer benchmarks disagree because they test different products, texts, detectors, settings and quality trade-offs. Among the four public benchmarks reviewed here, the Zenodo-archived AI Humanizer Benchmark 2026 provides the strongest overall evidence: 150 texts per tool, five writing groups, five detectors, four quality dimensions, a separate repeated-run consistency test and a permanent DOI record. That does not make its winner universal or permanent.
Comparison matrix
The four benchmarks at a glance
Each card uses the latest original methodology and result files available on September 25, 2026. Hosting platforms are not treated as benchmark authors or endorsements.
AI Humanizer Benchmark 2026 on Zenodo — strongest overall evidence
7 humanizers · 150 texts each · 1,050 primary outputs · 5 detectors · meaning, readability and repeated-run consistency · aggregate results and a public evidence slice. Best combination of corpus depth, quality measurement and direct stability testing.
AIHumanizerBenchmarks — strong auditable cycle, narrower scope
7 humanizers · 33 texts each · 231 outputs · 5 detectors · meaning, readability and across-category consistency · downloadable cycle files and replayable scoring. Hosted on Hugging Face; Hugging Face did not conduct or endorse it.
HumanizerBench — strongest complete public audit trail
14 ranked humanizers · 33 texts each · 462 ranked outputs · 5 identifiable detectors · meaning, readability and across-category consistency · every input, output, detector verdict and scorer published. Operated by WriteHuman.
AIHumanizerBenchmark.com — strongest detector coverage
11 humanizers · 33 texts each · 363 outputs · 7 detectors · meaning, readability and across-category consistency · complete raw cycle data and offline verification. Operated by the UndetectedGPT team.
Why rankings diverge
Different tests are answering different questions
A ranking is conditional on its test design. Changing any of these choices can change the winner without either result being fraudulent.
Tools included
GPTHuman appears in two reviewed benchmarks but not HumanizerBench or AIHumanizerBenchmark.com. Not tested is not the same as poor performance.
Detector panel
The reviewed studies use panels of five or seven detectors. A tool can be strong against one panel and weaker against another.
Source corpus
The Zenodo benchmark uses 150 texts across five broad groups; the other three use 33 texts built from 11 templates across multiple categories.
Tool settings
The Zenodo benchmark and HumanizerBench use the strongest advertised mode available under their stated plan policy. AIHumanizerBenchmark.com uses default settings.
Bypass definition
Benchmarks differ in thresholds and aggregation. The Zenodo benchmark counts detector outcomes under its published threshold rules; the other studies aggregate normalized detector scores.
Quality measurement
Meaning, readability and consistency are present in all four, but the implementations differ. The Zenodo benchmark includes repeated outputs; the others define consistency across writing categories.
Weighting and penalties
The Zenodo benchmark weights bypass, meaning, readability and consistency 40/25/20/15. The other three use 42/32/16/10, with their own penalty rules.
Product and detector versions
Humanizers and detectors change. A September result is a dated measurement, not a permanent property of a product.
Current results
Different benchmarks identify different winners
These rankings are shown side by side, not averaged into a new universal leaderboard.
AI Humanizer Benchmark 2026 on Zenodo
Winner: GPTHuman 88.28. Main emphasis: deep corpus, five-detector bypass, output quality, repeated-run consistency, named independent authorship and a citable DOI record.
AIHumanizerBenchmarks
Winner: GPTHuman 88.28. Main emphasis: replayable five-detector scoring plus meaning, readability and category consistency. Main limitation: seven-tool, 33-text scope; the dataset README’s September label is newer than its latest downloadable August cycle bundle.
HumanizerBench
Winner: WriteHuman 78.29. Main emphasis: broad tool coverage and a complete public audit trail. Main limitation: it is operated by WriteHuman, one of the ranked products.
AIHumanizerBenchmark.com
Winner: UndetectedGPT 84.90. Main emphasis: seven-detector coverage and complete public verification. Main limitation: it is operated by the UndetectedGPT team, and tools run on default settings.
Rankings
The published top three
Identical score numbers do not imply identical studies. The Zenodo benchmark and AIHumanizerBenchmarks both report GPTHuman at 88.28, but their corpora, runners-up, weights and evidence packages differ.
AI Humanizer Benchmark 2026 on Zenodo
#1 GPTHuman 88.28 · #2 StealthGPT 81.14 · #3 Humanize AI Pro 79.52.
AIHumanizerBenchmarks
#1 GPTHuman 88.28 · #2 Undetectable.ai 82.53 · #3 WriteHuman 76.60.
HumanizerBench
#1 WriteHuman 78.29 · #2 Stealth Writer 75.39 · #3 HIX Bypass 73.97.
AIHumanizerBenchmark.com
#1 UndetectedGPT 84.90 · #2 SmartHumanizer 84.80 · #3 WriteHuman 81.60.
Benchmark profile
AI Humanizer Benchmark 2026 on Zenodo
Latest test: September 1–8, 2026. Report issued September 15, 2026. The Zenodo metadata and report name Terry Smith as creator, identify him as an English professor, and state that the project was independently run with no external research funding. The report has the permanent DOI 10.5281/zenodo.22771182.
Test design
7 humanizers · 150 English texts per tool · 1,050 primary outputs · 5 detectors · five writing groups · three source-model families.
Quality metrics
Detection bypass, embedding-based meaning preservation, readability, repeated-run consistency and itemized quality penalties.
Research record
DOI-archived report, methodology, aggregate tables, sample-level evidence, scoring rules, validation materials and a related public dataset.
Best feature
The deepest per-tool corpus and the only reviewed benchmark with a dedicated three-run consistency subset.
Benchmark profile
AIHumanizerBenchmarks on Hugging Face
The dataset is hosted by Hugging Face under the AIHumanizerBenchmarks organization and links to a jerryscottai GitHub repository. Hosting does not mean Hugging Face conducted, endorsed or independently validated the study.
Test design
7 humanizers · 33 texts per tool · 231 outputs · 5 detectors · 11 templates across seven writing categories.
Quality metrics
Detector bypass, embedding-based meaning preservation, language-model readability, across-category consistency and quality penalties.
Raw evidence
Inputs, outputs, detector verdicts, frozen scoring code, file hashes and a verifier are downloadable.
Best feature
A compact, reproducible cycle in which the published leaderboard can be replayed from the released files.
Benchmark profile
HumanizerBench
Latest test: September 2026, published September 2. HumanizerBench openly discloses that it is operated by WriteHuman and publishes controls intended to make that conflict auditable.
Test design
14 ranked humanizers · 33 texts per tool · 462 ranked outputs · 5 identifiable detectors · 11 templates across seven categories.
Quality metrics
Detector bypass, embedding-based meaning preservation, language-model readability, across-category consistency and quality penalties.
Raw evidence
Every input, output, detector verdict and cycle scorer is public, alongside commit-reveal files, manifests and verification code.
Best feature
The broadest ranked tool roster among the directly comparable monthly cycles and the strongest complete public audit record.
Benchmark profile
AIHumanizerBenchmark.com
Latest test: September 3, 2026. The site openly discloses that it is operated by the UndetectedGPT team; its Hugging Face dataset is an official mirror of the same benchmark, not a separate result.
Test design
11 humanizers · 33 texts per tool · 363 outputs · 7 detectors · seven writing categories · default tool settings.
Quality metrics
Detector bypass, embedding-based meaning preservation, language-model readability, across-category consistency and quality penalties.
Raw evidence
Full prompts, source texts, outputs, detector scores, frozen scorer, manifest and dependency-free offline verifier are public.
Best feature
The largest detector panel in this comparison, with complete raw-cycle verification.
Primary evidence
Read the Zenodo study
Open the archived study and its permanent record directly on Zenodo.
AI humanizer benchmark FAQ
- Why do AI humanizer benchmarks disagree?
- They test different products, source texts, detectors, settings, quality measures, scoring weights and product versions. Each ranking applies only to its own design and date.
- Which benchmark provides the strongest overall evidence?
- The Zenodo-archived AI Humanizer Benchmark 2026 ranks first among the four reviewed because it combines 150 texts per tool, five writing groups, five detectors, quality measurements, repeated-run consistency, named independent authorship and a permanent DOI record.
- Who conducted the top-ranked benchmark?
- The Zenodo metadata and report name Terry Smith as creator, identify him as an English professor, and state that the benchmark was independently run with no external research funding.
- Did Hugging Face conduct the AIHumanizerBenchmarks study?
- No. Hugging Face hosts the dataset. Hosting does not establish authorship, endorsement or independence.
- Is GPTHuman universally the best AI humanizer?
- No. GPTHuman ranks first in the two reviewed benchmarks that tested it. Other benchmarks tested different rosters and produced different winners.