Updated November 18, 2026 November edition
AI humanizers.
Put to the test.
Open weights and closed services, measured on the same texts. HumanizeBench compares 23 entries across seven detectors, with meaning preservation and readability scored under one shared framework. Explore every source, rewrite, and measured result.
November 2026 AI Humanizer Benchmark
23 entries · 33 shared source texts per entry · 7 detectors
Measurement coverage5,252 / 5,313
Scores are provisional while 58 detector checks remain unresolved. Rankings update as measurements arrive. Provider exclusions: 3. Excluded checks have no numeric score. Readability and meaning: 759/759 recorded.
Inspect coverage and recorded evidence →| Rank ⓘ | Humanizer | Overall ⓘ ↓ | Bypass ⓘ | Meaning ⓘ | Readability ⓘ | Penalties ⓘ | Coverage | Evidence |
|---|---|---|---|---|---|---|---|---|
| 133 examples | 55.34 | 78.0 | 35.7 | 73.4 | △ −12.0 | 231/231 | ||
| 233 examples | 54.81 | 77.3 | 25.4 | 75.8 | △ −10.0 | 231/231 | ||
| 333 examples | 50.65 | 61.3 | 46.8 | 79.8 | △ −10.0 | 231/231 | ||
| 433 examples | Humanizer · CMFVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.infosave/humanizer-cmf | 45.23 | 30.2 | 65.9 | 78.0 | △ −3.0 | 231/231 | |
| 533 examples | Humanizer 12B · GGUFVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.jialinyyzz/humanizer-GGUF | 45.19 | 25.5 | 69.6 | 80.0 | △ −2.0 | 231/231 | |
| 633 examples | Humanizer 12BVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.jialinyyzz/humanizer | 45.19 | 25.5 | 69.6 | 80.0 | △ −2.0 | 231/231 | |
| 733 examples | Smaug HumanizerVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.abacusai/Smaug-Humanizer | 44.18 | 18.5 | 76.5 | 81.3 | △ −1.0 | 231/231 | |
| 833 examples | Humanizer 12B · OptiQ 4-bitVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.mlx-community/humanizer-12b-OptiQ-4bit | 43.13 | 36.3 | 60.9 | 77.1 | △ −7.0 | 231/231 | |
| 933 examples | Qwen 2.5 14B HumanizerVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.arunsingh80475/qwen2.5-14b-humanizer-v4-lora | 43.12 | 66.2 | 21.1 | 78.1 | △ −15.0 | 230/231 | |
| 1033 examples | Humanizer 12B · MLX 8-bit (cowWhySo)Verified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.cowWhySo/humanizer-12b-mlx-8bit | 41.93 | 25.3 | 66.5 | 79.0 | △ −4.0 | 230/231 | |
| 1133 examples | 41.12 | 46.5 | 43.2 | 74.6 | △ −10.0 | 231/231 | ||
| 1233 examples | Humanizer 12B · MLX 8-bit (npario)Verified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.npario/humanizer-12b-mlx-8bit | 40.34 | 24.6 | 65.9 | 78.8 | △ −5.0 | 229/231 | |
| 1333 examples | 36.44 | 38.6 | 46.4 | 79.9 | △ −12.0 | 231/231 | ||
| 1433 examples | 35.63 | 37.7 | 41.4 | 71.7 | △ −10.0 | 231/231 | ||
| 1533 examples | 34.05 | 29.0 | 50.2 | 77.6 | △ −10.0 | 231/231 | ||
| 1633 examples | GoHumanizeVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.gohumanize/gohumanize-open-humanizer | 32.90 | 0.2 | 92.0 | 85.1 | △ −7.0 | 231/231 | |
| 1733 examples | Open Humanizer 4BVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.akhaliq/open-humanizer-4b | 30.91 | 0.0 | 96.6 | 85.1 | △ −10.0 | 231/231 | |
| 1833 examples | 30.70 | 45.4 | 41.3 | 76.9 | △ −20.0 | 231/231 | ||
| 1933 examples | 26.21 | 0.8 | 60.2 | 82.2 | △ −6.0 | 231/231 | ||
| 2033 examples | Humanizer 1B · OptiQ 4-bitVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.mlx-community/humanizer-1B-OptiQ-4bit | 17.81 | 0.6 | 77.1 | 71.2 | △ −17.0 | 229/231 | |
| 2133 examples | 12.61 | 0.5 | 55.6 | 79.6 | △ −18.0 | 231/231 | ||
| 2233 examples | HumanizerAI · GGUFVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.mradermacher/humanizerAI-GGUF | 12.11 | 1.0 | 75.9 | 82.6 | △ −24.0 | 199/231 | |
| 2333 examples | Humanizer 3BVerified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.mondk/Humanizer-3B-Safetensors | 0.00 | 1.3 | 12.3 | 72.1 | △ −20.0 | 208/231 |
Gold, silver and bronze highlight the current top three overall values. All values use our measurements. Scores use available detector measurements. Coverage and methodology are available above.
No entries match these filters.
Go beyond the score
Read what the models actually wrote, then inspect the measurements behind it.
Academic essay
The repair-and-refurbish electronics movement deserves strong public support because it reduces waste, saves consumers money, and challeng…
Academic essay
## Introduction Community-supported fisheries (CSFs) have emerged as an alternative seafood distribution model designed to connect consum…
Application essay
Dear Hiring Manager and Selection Committee, I am writing to express my enthusiastic interest in the High School Chemistry Teacher positi…
Application essay
My interest in social psychology comes from a desire to understand how people’s thoughts, feelings and behaviour are shaped by their relat…
How scores are computed
Full methodologyBypass
Median detector human scores per output, averaged across the entry.
Meaning
Embedding cosine similarity, rescaled by the original scorer.
Readability
Clarity, fluency, and naturalness under one frozen NVIDIA rubric.
Consistency
How evenly detector results hold up across writing categories.
Penalties
Fixed deductions for output issues. Each rule is capped at 10 points; final scores stay between 0 and 100.
Identical output−2
Whitespace-normalized output matches its source.
Refusal−1
The fixed text rule identifies a refusal.
Meaning drift−1
Raw embedding cosine is below 0.85.
Length inflation−1
The output exceeds 1.4× the source word count.
Length deflation−1
The output is below 0.6× the source word count.
Why compare open and closed humanizers?
A service name and a downloadable checkpoint tell you very little about the resulting text. We wanted to place their recorded outputs under one evaluation framework, with enough evidence to see where the numbers come from.
We keep source texts fixed, score the saved rewrites, and expose coverage gaps. Related quantizations remain visible as related entries. Approximate measurements remain labeled.
Why we built HumanizeBenchOne test set.
One scoring framework.
The run starts with 33 published source texts, not newly generated answers to similar prompts. Thirteen open-weight repository entries each produce one saved rewrite per source, joining 330 existing rewrites from ten closed-source services.
All 759 outputs share seven detector routes, one readability rubric, and one meaning-preservation profile. HumanizeBench v2 combines the measurements with additional weight on bypass. Read the complete method.
Find the right view for your use case
Explore the same recorded outputs by category, detector, or model type. No new examples are generated for these views.
Best for overall
Balanced results across every source category.
Best for students
Academic-essay category scores from the same recorded outputs.
Best for essays
Academic and application essay categories together.
Best for seo content
Blog-post and marketing-copy categories.
Best for academic writing
Writing quality and category results for the essay samples.
Best for marketing
Marketing-copy and blog-post category scores.
Best for open-weight models
Only the 13 locally or cloud-executed repository entries.
Best for meaning preservation
Semantic preservation, without using detector scores.
Seven detectors.
Every result visible.
Explore reported human-likelihood scores, collection coverage, and the exact rewrites behind each detector’s results.
GPTZero
702 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
ZeroGPT
759 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
Copyleaks
759 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
Originality.ai
756 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
Pangram
758 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
Grammarly
759 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
QuillBot
759 of 759 measurements saved · mean human-likelihood scores and entry-level evidence.
Frequently asked
questions
What the numbers mean, where the outputs came from, and what this benchmark can tell you.
Is there a final winner yet?+
The leaderboard is provisional while detector requests remain unresolved. A finished collection can include explicit provider exclusions, which are shown as unscored and handled as null by the original scorer. Numeric coverage remains visible for every entry.
Are open and closed entries scored the same way?+
Yes. All saved rewrites use the same detector routes, NVIDIA readability rubric, embedding profile, classification rules, weights, and penalties. Closed-source rewrites are published historical outputs; open-weight rewrites are from our recorded GPU runs.
What does “meaning” measure?+
It starts as cosine similarity between source and output embeddings. The original scorer rescales the interval from 0.75 to 1.00 onto 0 to 100. These original thresholds are retained without calibration for our NVIDIA embedding substitute.
Why are there several Humanizer 12B entries?+
The requested roster includes different formats, conversions, and repositories. They remain separate tested entries; related weight families are identified in the model pages and should not be counted as independent model designs.
Does a high detector score prove text is human-written?+
No. It is a measurement of the selected detector response on that output. It is not proof of authorship, and this benchmark makes no claim of guaranteed bypass on unseen text.
Can I inspect the actual rewrites?+
Yes. The evidence pages contain all 33 source texts, the 759 saved rewrites, measured quality scores, and available detector scores. JSON downloads are available without an account.
Is this identical to the original evaluation?+
No. The source texts and scoring arithmetic are preserved, but we use the agreed NVIDIA readability and embedding substitutes, a reconstructed rubric, current detector routes, and a shared conservative classification heuristic.
Notes from our benchmark
Original research notes about this run, its measurements, and its limits.
Methodology · 5 min read
One rubric for open weights and closed services
Why we re-score saved outputs instead of mixing historical and fresh quality ratings.
Read our findingsModel execution · 4 min read
A different file is not always a different model
Thirteen repository entries, related weight families, and what the comparison actually measures.
Read our findingsResearch note · 4 min read
What we can measure before the detectors finish
The complete readability and meaning results, and the limits of a detector-free comparison.
Read our findings