Dataset Viewer
Auto-converted to Parquet Duplicate
cycle
stringdate
2026-09-01 00:00:00
2026-11-01 00:00:00
run_date
stringclasses
3 values
rank
int64
1
11
humanizer
stringlengths
6
15
humanizer_name
stringlengths
6
15
composite
float64
33.6
88.1
bypass_rate
float64
0
0.98
meaning_preservation
float64
0.53
0.9
readability
float64
0.64
0.84
consistency
float64
0.68
1
penalty_total
int64
0
13
tests_completed
int64
33
33
settings_used
stringlengths
13
74
plan_tier_used
stringclasses
6 values
detector_copyleaks
float64
0
0.94
detector_gptzero
float64
0
1
detector_grammarly
float64
0.19
0.97
detector_originality
float64
0
1
detector_pangram
float64
0
0.17
⌀
detector_quillbot
float64
0.13
0.94
detector_winston
float64
0
0.95
⌀
detector_zerogpt
float64
0.26
0.95
methodology_version
stringclasses
3 values
scoring_version
stringclasses
2 values
prompt_set_version
stringclasses
1 value
category_academic_essay
float64
38.1
90.5
category_application_essay
float64
42
88.1
category_blog_post
float64
46.2
87.9
category_business_email
float64
40.2
84.3
category_discussion_board
float64
38.9
91.7
category_marketing_copy
float64
43.1
89.2
category_news_article
float64
43.4
92.4
2026-11
2026-11-01T21:02:21.250Z
1
cleverhumanizer
CleverHumanizer
79.33
0.8957
0.6368
0.7948
0.862
0
33
style=casual,type_generation=humanize
default
0.8788
0.8309
0.9685
0.3109
0.0876
0.9039
null
0.8106
1.2.0
1.0.1
1.0.0
81
68.1
81.2
78.6
79.8
84.8
84
2026-11
2026-11-01T21:02:21.250Z
2
undetectedgpt
UndetectedGPT
76.03
0.7069
0.7902
0.8373
0.7654
0
33
ultra_stealth=off, model=ghost-2
default
0.7576
0.6121
0.9306
0.0773
0.0618
0.9233
null
0.7133
1.2.0
1.0.1
1.0.0
82.2
67.7
69.3
60
83.5
83.5
87.4
2026-11
2026-11-01T21:02:21.250Z
3
rephrasy
Rephrasy
72.97
0.7996
0.696
0.7527
0.8073
3
33
model=v4,words=true
default
0.8788
0.8258
0.8645
0.2806
0.1718
0.8394
null
0.7553
1.2.0
1.0.1
1.0.0
78.4
61.4
79
70.2
84.7
82.3
78.7
2026-11
2026-11-01T21:02:21.250Z
4
gptinf
GPTinf
72.05
0.6984
0.715
0.7839
0.7299
0
33
style=standard,alg=0
default
0.6667
0.5412
0.8748
0.0624
0
0.8545
null
0.8105
1.2.0
1.0.1
1.0.0
83.4
63.6
66.2
54.6
82.7
73.2
82.5
2026-11
2026-11-01T21:02:21.250Z
5
superhumanizer
SuperHumanizer
69.41
0.5524
0.7894
0.7827
0.8422
0
33
mode=transform,tone=default,preserveHeadings=true
default
0.6061
0.3706
0.8582
0.107
0.053
0.8758
null
0.6459
1.2.0
1.0.1
1.0.0
76
75.5
70.4
59.2
68.4
65.6
60.9
2026-11
2026-11-01T21:02:21.250Z
6
smarthumanizer
SmartHumanizer
67.41
0.487
0.8263
0.8161
0.7454
0
33
tone=balanced
default
0.7576
0.2485
0.7648
0.043
0
0.7982
null
0.6198
1.2.0
1.0.1
1.0.0
79.4
63.1
63.5
50.8
80.1
61.1
76.4
2026-11
2026-11-01T21:02:21.250Z
7
hix-bypass
HIX Bypass
60.04
0.5603
0.8123
0.81
0.755
10
33
modelType=latest,source=web-tools
default
0.6718
0.4424
0.8661
0.0464
0
0.7918
null
0.7488
1.2.0
1.0.1
1.0.0
67.6
70.9
66.6
51
83.3
76.4
73.2
2026-11
2026-11-01T21:02:21.250Z
8
stealthgpt
StealthGPT
52.87
0.382
0.6511
0.8155
0.7944
5
33
tone=College,mode=Medium,model=heavy,qualityMode=quality,writingMode=essay
default
0.3939
0.3867
0.6512
0.0909
0.1685
0.56
null
0.5269
1.2.0
1.0.1
1.0.0
69
46.8
56.1
54.3
50.3
60.2
67.9
2026-11
2026-11-01T21:02:21.250Z
9
rehumanize
ReHumanize
44.65
0.033
0.6978
0.8221
0.978
2
33
length=standard,tone=normal,purpose=general_writing,model=basic1
default
0
0.01
0.5882
0.0433
0
0.2715
null
0.3368
1.2.0
1.0.1
1.0.0
43.6
47.9
46.2
50.5
49.1
46.6
44.9
2026-11
2026-11-01T21:02:21.250Z
10
undetectable-ai
Undetectable AI
33.71
0.0348
0.7328
0.7652
0.9557
13
33
readability=University,purpose=Essay,strength=Balanced
default
0
0.0297
0.2712
0.0364
0
0.1267
null
0.4652
1.2.0
1.0.1
1.0.0
39.3
45.4
52.7
49.8
41
51
46.3
2026-10
2026-10-01T05:05:32.822Z
1
undetectedgpt
UndetectedGPT
88.12
0.9601
0.7737
0.8364
0.966
0
33
ultra_stealth=off, model=ghost-2
default
0.7879
0.7195
0.9682
0.9665
null
0.9098
0.8439
0.9226
1.1.0
1.0.1
1.0.0
90.5
88.1
87.9
84.3
91.7
84
92.4
2026-10
2026-10-01T05:05:32.822Z
2
hix-bypass
HIX Bypass
83.87
0.9347
0.7791
0.643
0.9387
0
33
modelType=latest,source=web-tools
default
0.8182
0.0402
0.8888
0.8907
null
0.9342
0.8661
0.9472
1.1.0
1.0.1
1.0.0
86.4
80.2
83
79.8
89
84.1
86.4
2026-10
2026-10-01T05:05:32.822Z
3
gptinf
GPTinf
80.74
0.8866
0.6835
0.7767
0.9206
0
33
model=free1,style=standard,alg=0
default
0.5758
0.5511
0.8845
0.9215
null
0.8642
0.8915
0.8447
1.1.0
1.0.1
1.0.0
81.9
79.7
77.6
79.2
84
81.1
84.1
2026-10
2026-10-01T05:05:32.822Z
4
smarthumanizer
SmartHumanizer
80.2
0.7818
0.8294
0.7745
0.8427
0
33
tone=balanced
default
0.7273
0.2828
0.8158
0.8293
null
0.804
0.6
0.7709
1.1.0
1.0.1
1.0.0
87.6
75.6
72.2
74.5
88.6
81.4
85.4
2026-10
2026-10-01T05:05:32.822Z
5
aihumanize
AI Humanize
78.33
0.8955
0.6643
0.7209
0.893
1
33
auto=1,cjtype=0,model=1
default
0.7273
0.4093
0.9109
0.8546
null
0.8801
0.8652
0.8566
1.1.0
1.0.1
1.0.0
84.2
73.5
78.5
77.2
83.1
79.2
81.6
2026-10
2026-10-01T05:05:32.822Z
6
superhumanizer
SuperHumanizer
78.23
0.8392
0.7518
0.7103
0.8556
1
33
mode=transform,tone=default,preserveHeadings=true
default
0.6364
0.4758
0.8352
0.8824
null
0.8126
0.6882
0.8349
1.1.0
1.0.1
1.0.0
82.6
76.8
76.9
66.6
85.7
80.6
85.3
2026-10
2026-10-01T05:05:32.822Z
7
writehuman
WriteHuman
75.06
0.9762
0.5324
0.7139
0.96
4
33
tone=standard,enhanced=false
default
0.9394
0.9996
0.9594
0.9979
null
0.9251
0.9485
0.9271
1.1.0
1.0.1
1.0.0
81.1
79.8
77.2
73.9
81.3
77.7
82.8
2026-10
2026-10-01T05:05:32.822Z
8
cleverhumanizer
CleverHumanizer
68.06
0.8589
0.655
0.7682
0.7734
9
33
style=casual,type_generation=humanize
default
0.697
0.7835
0.8821
0.9384
null
0.7692
0.8515
0.781
1.1.0
1.0.1
1.0.0
79.7
70.2
85.1
62.3
79.3
77.1
81.8
2026-10
2026-10-01T05:05:32.822Z
9
stealthgpt
StealthGPT
64.52
0.6277
0.6771
0.8064
0.7588
4
33
tone=College,mode=Medium,model=heavy,qualityMode=quality,writingMode=essay
default
0.6061
0.4149
0.6897
0.6269
null
0.6129
0.5527
0.6741
1.1.0
1.0.1
1.0.0
78.3
51.2
70.1
59.5
69.6
77.3
70.8
2026-10
2026-10-01T05:05:32.822Z
10
rehumanize
ReHumanize
41.6
0.0165
0.6868
0.8273
0.9691
4
33
length=standard,tone=normal,purpose=general_writing,model=basic1
default
0.0471
0
0.5021
0.0003
null
0.261
0.0003
0.4808
1.1.0
1.0.1
1.0.0
42.1
46.1
48.6
47.9
46.3
43.1
47.5
2026-10
2026-10-01T05:05:32.822Z
11
undetectable-ai
Undetectable AI
33.6
0.0182
0.7294
0.7324
0.978
12
33
readability=University,purpose=Essay,strength=Balanced
default
0.0154
0.1202
0.2061
0.0381
null
0.1446
0.0297
0.3452
1.1.0
1.0.1
1.0.0
39.6
44.5
53
49
39
43.9
51.7
2026-09
2026-09-03T16:11:14.775Z
1
undetectedgpt
UndetectedGPT
84.9
0.8626
0.8476
0.7733
0.9146
0
33
ultra_stealth=off, model=ghost-2
base
0.2121
0.8001
0.8933
0.7918
null
0.9265
0.5445
0.8508
1.0.0
1.0.0
1.0.0
89.2
86.2
77.6
79.7
90.3
85.8
85.9
2026-09
2026-09-03T16:11:14.775Z
2
smarthumanizer
SmartHumanizer
84.8
0.8528
0.876
0.7573
0.8826
0
33
tone=balanced
base
0.1515
0.71
0.8691
0.7818
null
0.9423
0.51
0.8262
1.0.0
1.0.0
1.0.0
89.8
83.3
82.4
75.6
88.3
89.2
79.3
2026-09
2026-09-03T16:11:14.775Z
3
writehuman
WriteHuman
81.6
0.9038
0.7618
0.7236
0.8642
1
33
tone=standard,enhanced=false
free-anonymous
0.6364
0.5648
0.9418
0.5916
null
0.9224
0.8212
0.9195
1.0.0
1.0.0
1.0.0
86.2
84.9
80.6
69.5
83.1
86.2
79.6
2026-09
2026-09-03T16:11:14.775Z
4
gptinf
GPTinf
78.3
0.7986
0.757
0.767
0.8222
0
33
model=free1,style=standard,alg=0
free-anonymous
0.5758
0.698
0.8761
0.3378
null
0.849
0.6297
0.8517
1.0.0
1.0.0
1.0.0
85.3
79.5
75
61.8
82.9
81.7
73.2
2026-09
2026-09-03T16:11:14.775Z
5
aihumanize
AI Humanize
76.1
0.8263
0.7077
0.747
0.6793
0
33
auto=1,cjtype=0,model=1
free-anonymous (WEB_EMUSER)
0.6667
0.5536
0.8839
0.2321
null
0.8925
0.8103
0.8711
1.0.0
1.0.0
1.0.0
83.1
74.7
79.7
41.5
73
82.7
82.2
2026-09
2026-09-03T16:11:14.775Z
6
hix-bypass
HIX Bypass
73.8
0.7317
0.8082
0.6376
0.6989
0
33
modelType=latest,source=web-tools
free-anonymous
0.5455
0.1573
0.88
0.2462
null
0.889
0.6906
0.86
1.0.0
1.0.0
1.0.0
83.4
71
73.3
40.2
69.5
80.4
85.7
2026-09
2026-09-03T16:11:14.775Z
7
stealthgpt
StealthGPT
71.9
0.7463
0.7012
0.8267
0.7847
3
33
tone=College,mode=Medium,model=heavy,qualityMode=quality,writingMode=essay
paid-api
0.5758
0.7823
0.7755
0.4818
null
0.6598
0.6209
0.6648
1.0.0
1.0.0
1.0.0
80
66.5
76.6
81.3
57.2
85
68.8
2026-09
2026-09-03T16:11:14.775Z
8
superhumanizer
SuperHumanizer
68.5
0.707
0.773
0.7403
0.7276
5
33
mode=transform,tone=default,preserveHeadings=true
free-anonymous
0.5137
0.64
0.7682
0.5081
null
0.7456
0.5227
0.7871
1.0.0
1.0.0
1.0.0
82.4
64.5
68.1
57.1
81.3
77.6
85.5
2026-09
2026-09-03T16:11:14.775Z
9
cleverhumanizer
CleverHumanizer
66.8
0.865
0.6534
0.7321
0.8883
11
33
style=casual,type_generation=humanize
free-guest
0.5317
0.8999
0.9688
0.6701
null
0.8882
0.7045
0.7791
1.0.0
1.0.0
1.0.0
81.4
71.7
81.5
68.3
77.6
79.1
82.9
2026-09
2026-09-03T16:11:14.775Z
10
rehumanize
ReHumanize
39
0.037
0.5915
0.8142
0.9513
4
33
length=standard,tone=normal,purpose=general_writing,model=basic1
free-anonymous
0.0303
0.0573
0.6424
0.0006
null
0.219
0
0.3256
1.0.0
1.0.0
1.0.0
38.1
42
47.2
47.1
38.9
44.6
43.4
2026-09
2026-09-03T16:11:14.775Z
11
undetectable-ai
Undetectable AI
38.2
0.0024
0.8961
0.717
0.996
12
33
readability=University,purpose=Essay,strength=Balanced
paid-api
0
0.0611
0.1852
0
null
0.1657
0.0015
0.2648
1.0.0
1.0.0
1.0.0
48
47.8
52.9
50.9
50.8
50.9
51.6

AI Humanizer Benchmark: AI humanizers tested against 7 AI detectors (November 2026)

AI Humanizer Benchmark measures how well AI humanizers rewrite AI-generated text so that AI detectors classify it as human-written, and how much meaning and readability the rewrite loses. In each monthly cycle, the humanizers on the roster (10 in November 2026) rewrite the same 33 source texts across 7 writing categories, using each tool's default settings. Each output gets three kinds of score: 7 AI detectors (GPTZero, Originality.ai, Copyleaks, Pangram, ZeroGPT, QuillBot, Grammarly; Pangram replaced Winston AI from November 2026, and each cycle's leaderboard.json lists the panel it was scored with), meaning preservation (embedding similarity to the input) and readability (a language model's rating). From October 2026, the source texts are written by six language models from OpenAI, Anthropic and Google, and each row records which model wrote its text.

Before any tool runs, the test set is locked with a commit–reveal hash. Each cycle's scoring code is frozen and published with it, so the ranking below can be recomputed from this data. The dataset currently holds 1,056 humanizer outputs and 7,392 detector-score rows from three cycles (September to November 2026).

This dataset is the official mirror of the GitHub data repository, published by the AI Humanizer Benchmark organization. The website, the GitHub repo, and this dataset publish the same numbers; each links to the others.

Best AI humanizers: November 2026 ranking

# AI humanizer Score / 100 Bypass Meaning Readability Penalty
1 Clever AI Humanizer 79.33 90% 64% 79% 0
2 UndetectedGPT 76.03 71% 79% 84% 0
3 Rephrasy 72.97 80% 70% 75% −3.0
4 GPTinf 72.05 70% 72% 78% 0
5 SuperHumanizer 69.41 55% 79% 78% 0
6 SmartHumanizer 67.41 49% 83% 82% 0
7 HIX Bypass 60.04 56% 81% 81% −10.0
8 StealthGPT 52.87 38% 65% 82% −5.0
9 ReHumanize 44.65 3% 70% 82% −2.0
10 Undetectable AI 33.71 3% 73% 77% −13.0

Penalty is the total deducted for quality failures (severe meaning drift, length inflation or deflation, refusals, or output returned unchanged) and is already reflected in the score. Per-detector and per-category breakdowns for every tool: aihumanizerbenchmark.com/leaderboard. The leaderboard config below carries every published cycle as rows.

November 2026 runs under methodology 1.2.0. Each row carries its cycle's version stamps, and CHANGES.md records what moved and why.

The overall score is a 0–100 composite: detector bypass 42%, meaning preservation 32%, readability 16%, consistency across writing categories 10%, minus penalties for quality failures (meaning drift, length distortion, refusals). Full formula: methodology.

Contents

One directory per monthly cycle under data/cycles/<YYYY-MM>/:

  • leaderboard.json — every tool's scores: composite, bypass rate, meaning preservation, readability, consistency, per-detector and per-category breakdowns, penalties, and the exact settings used; from 2026-11 it also names the readability and embedding models
  • samples.json / prompts.json / templates.json / banks.json — the cycle's prompts: the source texts every tool rewrote, and how they were generated
  • source-models.json — the models that write the source texts, in the order published with the commitment (from 2026-10)
  • tests.json / results.jsonl — per-test rows: one entry per (tool, prompt) with meaning-preservation and readability values
  • detector-scores.json — every detector verdict for every output
  • data/*.jsonl — per-tool and per-detector files: the same rows grouped by tool or by detector; from 2026-11 each row also carries its cycle id, and per-tool rows their settings and detector verdicts
  • commit.json / nonce.txt / manifest.json — the commit–reveal record and file hashes that lock the test set before any tool runs. From 2026-10, commit.json also carries the list of source models
  • scoring.js — the frozen scoring code that produced this cycle's leaderboard

data/humanizers/*.json carries each tool's cross-cycle history. Published cycles are never edited; new cycles are added monthly.

Flat tables (load_dataset-ready)

For convenience, the same data is also published as flat row tables under flat/, derived mechanically from the canonical files above at publish time and never hand-edited:

  • tests — one row per (cycle, tool, prompt): the prompt, the AI-written input and the model that wrote it (source_model), the tool's output, per-detector human-probability verdicts (detector_* columns, 0 = judged AI, 1 = judged human), meaning preservation, and readability
  • leaderboard (default config) — one row per (cycle, tool): rank, composite, all sub-scores, penalties, the exact settings used, the cycle's version stamps, and the score in each writing category (category_* columns)
  • samples — one row per (cycle, prompt): each source text once, with its prompt, template, category and source_model
  • detector_scores — one row per (cycle, tool, prompt, detector): the same verdicts as the detector_* columns in long form, for filtering or grouping by detector
from datasets import load_dataset
tests = load_dataset("ai-humanizer-benchmark/ai-humanizer-benchmark", "tests")
leaderboard = load_dataset("ai-humanizer-benchmark/ai-humanizer-benchmark", "leaderboard")
samples = load_dataset("ai-humanizer-benchmark/ai-humanizer-benchmark", "samples")
detector_scores = load_dataset("ai-humanizer-benchmark/ai-humanizer-benchmark", "detector_scores")

Reading the columns:

  • cycle is on every row of every table. Filter on it to work with a single cycle.
  • source_model is the exact id of the model that wrote the source text. It is null for 2026-09, when a single model wrote every text.
  • detector_* and human_prob are the detector's human-probability for the tool's output. A null means there is no score for that test; detector_scores.outcome says why: score, no_verdict (the detector gives none for that text) or error (no verdict could be obtained).
  • meaning_preservation in tests is the raw embedding similarity between input and output. In leaderboard it is the tool's mean after rescaling, as used in the composite.
  • bypass_rate is the mean, over a tool's tests, of the median human-probability across the detector panel.
  • category_* is the composite computed on one writing category alone, before penalties.

New cycles append rows with their cycle value, and the values in existing rows are never modified. A column introduced by a later cycle is added to earlier rows as null. The canonical record remains the per-cycle files: any divergence between the flat tables and a recompute from those files would itself be a reportable defect.

Verifying the results

The scoring code is frozen per cycle and published (MIT). Recomputing the entire leaderboard from this raw data is one command in the GitHub repository:

npm run verify

Each cycle's prompt set is locked by a commit–reveal scheme (commit.json / nonce.txt): the nonce's SHA-256 is published before any tool runs, so prompts cannot be reselected after seeing which tools they favour. From 2026-10 the same nonce also decides which model writes each text, and the list of models is published with the commitment. Spot-checking a single verdict needs no code: paste any published output into the detector and compare. Every tool runs through this identical pipeline, including UndetectedGPT, whose team operates the benchmark (how conflicts are handled); the verification design exists precisely so no one has to trust the operator.

Citation

To cite the data as a whole, which always resolves to the newest published cycle:

AI Humanizer Benchmark. (2026). AI Humanizer Benchmark: monthly cycle data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22966275

Scores change from cycle to cycle, so a citation that quotes a specific number should reference that cycle's own DOI, listed under Versions on the Zenodo record.

Links

License

Cycle data: CC BY 4.0. Scoring code (in the GitHub repository): MIT.

Downloads last month
843