gpthuman/humanized-text-with-statistics

GPTHuman Humanizer Benchmark

This is a public benchmark of the GPTHuman humanizer. We built it to find out how good our humanizer truly is. We wanted to answer two questions honestly, not with hand-picked demos:

1. Does it get past AI detectors? Can GPTHuman rewrite AI-generated text so that leading commercial detectors classify it as human?
2. Does it keep the text intact? Does the rewrite keep the original meaning and facts, without glitches or broken text?

We took 500 AI-written English texts from public AI-detection datasets, humanized every one of them with GPTHuman, and scored each result with 4 commercial AI detectors and an LLM quality judge. Every row is published here, including the failures, so anyone can check the results.

Key findings

• 92.6% of texts got past all 4 external detectors (463 of 500).
• 92.4% passed every quality check (462 of 500): no glitches, meaning preserved, and facts preserved.
• The average similarity between the original and humanized text is 92.74 / 100.
• On average, text length changed by +14.2% in words and -5.8% in characters after humanization.

How the benchmark was built

1. Source texts. 500 texts written by 8 AI models across 10 content categories, sampled from the DetectRL-X and uChicago AI detection datasets. Every source text is AI-written (label = ai).
2. Humanization. Each text was rewritten once with GPTHuman, with no retries and no cherry-picking.
3. AI detection. Each humanized text was scored by ZeroGPT, Originality.ai, QuillBot, and Copyleaks. Every score is on the same 0-100 human scale: 100 means purely human, 0 means purely AI. A text counts as classified human when its score is 50 or higher.
4. Quality judging. An LLM compared each original with its humanized version and returned three pass/fail scores: hallucination, meaning preserved, and factual preserved.

AI detection results

Because every source text is AI-written, the bypass rate is the share of humanized texts a detector classified as human. Higher is better for the humanizer.

Detector	Mean	Median	Std dev	Min	25th pct	75th pct	Max	Bypass rate
ZeroGPT	99.49	100.0	3.26	41.6	100.0	100.0	100.0	99.8%
Originality.ai	93.02	100.0	22.07	0.1	99.8	100.0	100.0	93.2%
Copyleaks	99.4	100.0	7.72	0.0	100.0	100.0	100.0	99.4%
QuillBot	99.43	100.0	3.58	65.5	100.0	100.0	100.0	100.0%
GPTHuman (built-in)	98.59	98.7	1.67	71.0	98.5	98.9	99.9	100.0%

How many of the 4 external detectors each text got past:

Detectors bypassed	Texts	Share
4 of 4	463	92.6%
3 of 4	36	7.2%
2 of 4	1	0.2%
1 of 4	0	0.0%
0 of 4	0	0.0%

Quality results

Check	What it measures	Pass	Pass rate	Mean
No hallucination (hallucination)	No glitch such as doubled words or garbled text	497	99.4%	0.006
Meaning preserved (meaning_preserved)	Same main point, stance, and conclusion	500	100.0%	1
Facts preserved (factual_preserved)	No key fact changed	465	93.0%	0.93

hallucination is 1 when the humanized text contains a generation glitch: a doubled word ("I like like dogs"), characters from another writing system, random or garbled words, or leftover markup. Rewording is never counted as a hallucination. Its mean is the share of texts with a glitch, so lower is better.

Similarity and readability

Metric	Mean	Median	Std dev	Min	25th pct	75th pct	Max
similarity	92.74	94.1	3.2	81.6	90.62	95.3	96.8
readability	56.29	56.35	12.44	20.8	47.1	65.3	91.9

similarity measures how close the humanized text is to the original (0-100). readability is a reading-ease score for the humanized text, where higher means easier to read.

Length structure

How the length of each text changed between the original and the humanized version. Words are counted by splitting on whitespace, and characters exclude leading and trailing whitespace. Characters per word shows whether the rewrite uses shorter or longer words.

Measure	Mean	Median	Min	Max
Original words	317.82	285.0	143	637
Humanized words	351.56	323.0	173	713
Word difference (humanized - original)	+33.7	+39.0	-142.0	+126.0
Word difference (%)	+14.2	+13.4	-28.3	+62.7
Original characters	2169.65	1987.5	847	5157
Humanized characters	1978.79	1840.5	937	4247
Character difference (humanized - original)	-190.9	-115.5	-1286.0	+294.0
Character difference (%)	-5.8	-6.0	-37.1	+29.5
Original characters per word	6.82	6.79	5.5	8.69
Humanized characters per word	5.62	5.64	4.87	6.43

Change	Texts	Share
Shorter by more than 10% (words)	15	3.0%
Within 10% of the original (words)	185	37.0%
Longer by more than 10% (words)	300	60.0%

Dataset composition

Model	Texts
gemini-2.5-flash	80
deepseek-v3	78
qwen-max	65
gpt-4o	60
claude-sonnet-4	59
gemini-2.0-flash	55
claude-opus-4	54
gpt-4.1	49

Category	Texts
news	161
webtext	62
academic	62
seo	51
blog	45
wiki	44
resume	27
restaurant review	19
novel	17
amazon review	12

Columns

Column	Type	Description
id	int	Row id from the source dataset.
original	string	Original AI-generated text.
humanized	string	Text after GPTHuman humanization.
label	class	Source label. Always ai.
source	string	Source dataset.
category	string	Content type, for example news, academic, or blog.
model	string	AI model that wrote the original text.
similarity	float	Similarity between the original and humanized text, 0-100.
readability	float	Reading-ease score for the humanized text.
human_score	float	GPTHuman's built-in human score, 0-100.
zerogpt	float	ZeroGPT human score, 0-100.
originality	float	Originality.ai human score, 0-100.
copyleaks	float	Copyleaks human score, 0-100.
quillbot	float	QuillBot human score, 0-100.
hallucination	int	1 if the humanized text has a generation glitch, else 0.
meaning_preserved	int	1 if the main point, stance, and conclusion are unchanged.
factual_preserved	int	1 if no key fact (name, number, date, place, role, attribution) was changed.

Usage

from datasets import load_dataset

ds = load_dataset("gpthuman/humanized-text-with-statistics", split="train")

detectors = ["zerogpt", "originality", "copyleaks", "quillbot"]
bypassed_all = ds.filter(lambda r: all(r[d] >= 50 for d in detectors))
print(f"{len(bypassed_all)} / {len(ds)} texts bypassed every detector")

Limitations

• English, AI-written only. The benchmark has no human-written texts, so it doesn't measure false positives on real human writing.
• Point in time. Detector scores were collected on September 29, 2026. Detectors change often, so re-running them later may give different results.
• Fixed threshold. Bypass rates use a single cutoff of 50. Each detector's own interface may use different thresholds or labels such as "mixed".
• LLM judge. Quality scores come from an LLM, not human annotators. Rewording, style changes, and small rounding are not penalized.

License and sources

The original texts come from DetectRL-X and the uChicago AI detection dataset. Please follow the licenses of those datasets when reusing the original column.