Pangram AI share
The saved AI share for the humanized text, including manual scores. The main ranking gives each scored source model equal influence. A lower share does not prove that a person wrote the text. AI-assisted text is reported separately.
AI detection and stylistic analysis
AI Humanizer benchmark
Which AI humanizers actually change the writing? Compare their Pangram AI share, GPTZero results, and writing style across Claude Fable 5.1, GPT-6 Astra, GPT-4, and Gemini 3.1 Pro. Inspect the scores, then read the original and humanized text for yourself.
Compare the original LLMs in the Stylistic LLM Writing benchmark
Detailed Humanizer comparison
Lower Pangram AI shares rank higher. AI Style score measures writing patterns separately. Missing scores are not treated as zero.
Showing the last published results. Newer saved results are awaiting synchronization; detailed text comparisons are temporarily paused.
| Rank | Humanizer / model | Pangram AI shareRanks Humanizers. Lower is better. | GPTZero AI scoreAdditional detector score. Lower is better. | AI Style scoreStylistic evidence, separate from AI detection. | Sentence length diversityDistribution breadth, quantiles, and tails. | GPT-2 perplexity similaritySimilarity to model-like token predictability. | Noun contextualizationContext supplied around nouns. | Heavy grammarLLM-favored constructions per 1,000 words. | Personal toneVisible judgment and point of view. | AI SlopRepeated rhetorical templates per 1,000 words. | AI WordingLLM-overused words and three-word phrases per 1,000 words. | Punctuation richnessRare-sensitive variety across punctuation types. | Section variationDoes the style change from start to finish? |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| #1 | 0%Low AI | 0%Low AI | 16.2/100Low AI | 0.744Low AISentence lengths are broadly distributed. | 0.0139Low AIToken choices are within the lowest evidence quarter. | 0.853Low AINouns are well contextualized with linking, helping, and qualifying language. | 13.3/1kLow AIFew of the directional dense-grammar signals are elevated. | 7.3%Low AIPersonal-tone density is near the human anchor. | 6.5/1kLow AIFew repeated rhetorical templates. | 5.4/1kLow AIFew overused AI wording matches. | 0.766Low AIPunctuation is richly distributed across common and less-common mark types. | 0.31Low AIThe document changes shape naturally. | |
| #2 | 13.2%Low AI | 3.4%Low AI | 29.6/100Low AI | 0.683High AISentence lengths are concentrated in a few bands. | 0.0224Low AIToken choices are within the lowest evidence quarter. | 0.83Low AINouns are well contextualized with linking, helping, and qualifying language. | 24.5/1kLow AIFew of the directional dense-grammar signals are elevated. | 5.8%High AIVery little visible perspective or uncertainty. | 4.3/1kLow AIFew repeated rhetorical templates. | 8.4/1kLow AIFew overused AI wording matches. | 0.574High AIPunctuation is concentrated in relatively few mark types. | 0.254Medium AISome section-to-section change. | |
| #3 | 78.9%High AI | 19.5%Low AI | 33.5/100Low AI | 0.652High AISentence lengths are concentrated in a few bands. | 0.2421High AIToken choices are beyond the interval midpoint. | 0.829Low AINouns are well contextualized with linking, helping, and qualifying language. | 13.8/1kLow AIFew of the directional dense-grammar signals are elevated. | 6%Medium AISome personal or subjective language. | 4.5/1kLow AIFew repeated rhetorical templates. | 5.8/1kLow AIFew overused AI wording matches. | 0.584High AIPunctuation is concentrated in relatively few mark types. | 0.292Low AIThe document changes shape naturally. | |
| #4 | 85.4%High AI | 29%Low AI | 29.8/100Low AI | 0.762Low AISentence lengths are broadly distributed. | 0.2556High AIToken choices are beyond the interval midpoint. | 0.971Low AINouns are well contextualized with linking, helping, and qualifying language. | 24.9/1kLow AIFew of the directional dense-grammar signals are elevated. | 6%Medium AISome personal or subjective language. | 5.1/1kLow AIFew repeated rhetorical templates. | 6.7/1kLow AIFew overused AI wording matches. | 0.636Medium AISome less-common punctuation types are represented. | 0.233Medium AISome section-to-section change. | |
| #5 | 95.6%High AI | 43.4%Medium AI | 25.2/100Low AI | 0.704Medium AISentence lengths use several bands. | 0.0565Medium AIToken choices are between one quarter and one half of the interval. | 0.797Low AINouns are well contextualized with linking, helping, and qualifying language. | 14.6/1kLow AIFew of the directional dense-grammar signals are elevated. | 5.6%High AIVery little visible perspective or uncertainty. | 5.9/1kLow AIFew repeated rhetorical templates. | 7.8/1kLow AIFew overused AI wording matches. | 0.723Low AIPunctuation is richly distributed across common and less-common mark types. | 0.272Low AIThe document changes shape naturally. | |
| #6 | 97.8%High AI | 77%High AI | 30.3/100Low AI | 0.709Medium AISentence lengths use several bands. | 0.1633High AIToken choices are beyond the interval midpoint. | 0.924Low AINouns are well contextualized with linking, helping, and qualifying language. | 20.5/1kLow AIFew of the directional dense-grammar signals are elevated. | 5.4%High AIVery little visible perspective or uncertainty. | 3.8/1kLow AIFew repeated rhetorical templates. | 6.1/1kLow AIFew overused AI wording matches. | 0.643Medium AISome less-common punctuation types are represented. | 0.209High AIVery little style change across the document. | |
| #7 | 100%High AI | 82.3%High AI | 25.7/100Low AI | 0.721Medium AISentence lengths use several bands. | 0Low AIToken choices are within the lowest evidence quarter. | 0.854Low AINouns are well contextualized with linking, helping, and qualifying language. | 16.6/1kLow AIFew of the directional dense-grammar signals are elevated. | 6.1%Medium AISome personal or subjective language. | 9.3/1kLow AIFew repeated rhetorical templates. | 7.6/1kLow AIFew overused AI wording matches. | 0.71Low AIPunctuation is richly distributed across common and less-common mark types. | 0.221High AIVery little style change across the document. | |
| #8 | 100%High AI | 81.3%High AI | 28.5/100Low AI | 0.701Medium AISentence lengths use several bands. | 0.035Low AIToken choices are within the lowest evidence quarter. | 0.898Low AINouns are well contextualized with linking, helping, and qualifying language. | 14.8/1kLow AIFew of the directional dense-grammar signals are elevated. | 6%Medium AISome personal or subjective language. | 5.6/1kLow AIFew repeated rhetorical templates. | 6.7/1kLow AIFew overused AI wording matches. | 0.651Medium AISome less-common punctuation types are represented. | 0.209High AIVery little style change across the document. |
What the scores tell you
GPTZero AI score
An additional detector score for comparison. It does not determine the main ranking or change AI Style score.
AI Style score
A separate comparison of sentence lengths, noun contextualization, heavy grammar, personal tone, rhetorical templates, overused vocabulary, punctuation, section variation, and GPT-2 predictability.
Original versus humanized
Choose a source model to compare each Humanizer with the original. Open a row to inspect its individual writing tasks and read both versions.
Compare the same source
Use the source-model tab to compare how different Humanizers transform the same model’s writing. Open the examples to check whether each tool preserves the original meaning.
AI Humanizer benchmark FAQ
How are Humanizers ranked?
By their average Pangram AI share across scored source models, lower first. Each scored source model has equal influence. AI Style score breaks ties. Humanizers without Pangram scores appear after scored results.
Were the original models scanned by the detectors?
No. Original LLM rows use a 100% reference value and the human baseline uses 0% for both Pangram and GPTZero. These are reference values, not measured detector results. Detector scans apply to humanized outputs.
Does a lower score mean better writing?
Not necessarily. Detector scores and style measurements do not establish factual accuracy, usefulness, originality, or reader preference. Read the paired outputs to check whether the rewrite preserved the meaning.
Why compare different source models?
A Humanizer can perform differently on each model's style. The source-model tab separates these effects instead of hiding them in a single average.
How do the results stay up to date?
This page uses the benchmark owner's server-published results, including the same metric labels as the app. New tests appear after the updated table is synchronized. No tool is tested or charged when you view this page.
Can an AI humanizer guarantee bypassing detectors?
No. A result from one detector on one set of samples is not a guarantee for other texts, detector versions, or contexts.