Independent style analysis

LLM writing benchmark

Q: Does a high score prove that a text was written by AI?

No. The score compares measurable stylistic tendencies and is not proof of authorship for an isolated document.

Compare how frontier language models write across the same set of practical tasks. Lower AI evidence means the model output is stylistically closer to the human baseline used in this benchmark.

25 compared models and baseline 8 writing genres 6 equally weighted signals Updated August 6, 2026

Detailed model comparison

Scores update automatically when a new benchmark is published.

Low AI Medium AI High AI

Rank	AI ScoreAverage of six signals	Sentence rhythmSentence-length variation	GPT-2 proximityModel-like predictability	Syntax diversityGrammar-pattern variety	Prose heavinessDense constructions per 1,000 words	Personal tonePerspective signals per 100 words	Section variationIntro-body-ending change
#1	30.3/100Low AI	0.59Low AI	0.0056Low AI	0.891Low AI	102.8/1kLow AI	9.8%Low AI	0.297Low AI
#2	30.9/100Low AI	0.66Low AI	0.0001Low AI	0.885Medium AI	95.5/1kLow AI	5.9%High AI	0.242Low AI
#3	35.3/100Medium AI	0.64Low AI	0.0284Low AI	0.887Low AI	84/1kLow AI	4.7%High AI	0.244Low AI
#4	36.9/100Medium AI	0.56Low AI	0.0004Low AI	0.889Low AI	78.4/1kLow AI	4.9%High AI	0.235Medium AI
#5	39.9/100Medium AI	0.65Low AI	0Low AI	0.889Low AI	111/1kLow AI	7.1%High AI	0.209High AI
#6	41.5/100Medium AI	0.49High AI	0.0051Low AI	0.874High AI	104.1/1kLow AI	5.5%High AI	0.215High AI
#7	42.3/100Medium AI	0.55Low AI	0.0087Low AI	0.838High AI	73.7/1kLow AI	3.9%High AI	0.243Low AI
#8	43.1/100Medium AI	0.5High AI	0.007Low AI	0.884Medium AI	88.3/1kLow AI	5.7%High AI	0.197High AI
#9	43.3/100Medium AI	0.6Low AI	0.0145Low AI	0.887Low AI	109.1/1kLow AI	7.6%High AI	0.213High AI
#10	44.3/100Medium AI	0.65Low AI	0Low AI	0.88High AI	115.3/1kMedium AI	6.9%High AI	0.23Medium AI
#11	48.5/100Medium AI	0.52High AI	0.0033Low AI	0.882High AI	118.9/1kMedium AI	7.3%High AI	0.199High AI
#12	51/100High AI	0.53Medium AI	0.0021Low AI	0.856High AI	115.8/1kMedium AI	7.9%High AI	0.225High AI
#13	52.4/100High AI	0.48High AI	0.9996High AI	0.894Low AI	90.8/1kLow AI	4.8%High AI	0.293Low AI
#14	53.3/100High AI	0.49High AI	0.0212Low AI	0.876High AI	112.1/1kLow AI	5.7%High AI	0.17High AI
#15	54.1/100High AI	0.57Low AI	0.3335High AI	0.873High AI	85.1/1kLow AI	6.7%High AI	0.23Medium AI
#16	54.6/100High AI	0.51High AI	0.024Low AI	0.871High AI	104.9/1kLow AI	5.3%High AI	0.154High AI
#17	55.9/100High AI	0.49High AI	0.0157Low AI	0.874High AI	115.3/1kMedium AI	7.1%High AI	0.214High AI
#18	56.1/100High AI	0.5High AI	0.0003Low AI	0.866High AI	111/1kLow AI	5.7%High AI	0.218High AI
#19	58/100High AI	0.56Low AI	0.036Low AI	0.873High AI	123.9/1kHigh AI	7%High AI	0.2High AI
#20	59.5/100High AI	0.5High AI	0.0295Low AI	0.875High AI	105.7/1kLow AI	5%High AI	0.157High AI
#21	60.5/100High AI	0.42High AI	0.1528High AI	0.886Medium AI	122.1/1kHigh AI	7%High AI	0.238Low AI
#22	60.9/100High AI	0.51High AI	0.629High AI	0.875High AI	108/1kLow AI	10%Low AI	0.207High AI
#23	62.1/100High AI	0.57Low AI	0.1023High AI	0.866High AI	111.3/1kLow AI	10.6%Low AI	0.195High AI
#24	71.8/100High AI	0.43High AI	0.2718High AI	0.877High AI	130.4/1kHigh AI	7.9%High AI	0.143High AI
#25	73.4/100High AI	0.29High AI	0.9944High AI	0.893Low AI	135.4/1kHigh AI	7.6%High AI	0.213High AI

Metric labels: Low AI through 25% evidence, Medium through 50%, High above 50%. Global score: Low through 35, Medium through 50, High above 50.Updated August 6, 2026

How to read it

Style evidence, not authorship proof

The AI Score is the equal-weight average of six interpretable writing signals. Use it to compare model tendencies under controlled tasks, not to determine who wrote an isolated text.

1
Start with the AI Score
Lower scores rank higher. The score equally averages all six calibrated signals for every completed writing task.
2
Compare the metric labels
Green is near the low-evidence end of a calibration. Yellow is intermediate; red has crossed the interval midpoint.
3
Open a model, then a task
Model analytics show averages. Task analytics reveal genre-specific scores, graphs, recommendations and highlighted output.

Benchmark methodology

Six writing signals used in every AI Score

Every model is evaluated with the same metrics and equal weights. The final AI Score is the average of the six calibrated evidence values across that model's completed tasks.

Exact score calculation: task AI Score = 100 x mean(the six available 0-1 evidence values). Model AI Score = mean(all completed task AI Scores). Individual raw values are averaged only for display; the model score is not recalculated from those displayed averages.

Sentence rhythm

Counts the words in every sentence, calculates their population standard deviation, then divides it by the mean sentence length. A larger coefficient means a wider mix of short and long sentences.

Formula: SD(sentence word counts) / mean sentence word count.
Evidence anchors: 0.55 = 0; 0.50 = 1.

GPT-2 perplexity proximity

DistilGPT-2 compares the observed next-token log probability with the probability distribution it expected at every position. The standardized difference is converted through the normal cumulative distribution.

Formula: Phi((observed log probability - expected log probability) / sqrt(expected variance)).
Evidence anchors: 0.00 = 0; 0.20 = 1.

Syntax diversity

spaCy converts the text into part-of-speech sequences. The score combines normalized entropy for single tags, two-tag sequences, three-tag sequences, and unique sentence-level POS patterns.

Formula: 0.25 x unigram entropy + 0.35 x bigram entropy + 0.35 x trigram entropy + 0.05 x sentence-pattern diversity.
Evidence anchors: 0.890 = 0; 0.875 = 1.

Prose heaviness

Counts nominalizations, reduced participial clauses, subject-that constructions, and phrasal coordination, then normalizes their total for text length.

Formula: dense constructions / total words x 1,000.
Evidence anchors: 110/1k = 0; 130/1k = 1.

Personal tone

Counts personal pronouns, private-state verbs such as think or feel, hedges, and direct questions. It estimates visible perspective, uncertainty, and reader engagement.

Formula: (personal pronouns + private verbs + hedges + questions) / total words x 100.
Evidence anchors: 9% = 0; 7.5% = 1.

Section variation

Splits the text into opening, middle, and ending thirds. Each third is profiled by sentence length, sentence rhythm, function-word rate, and modifier rate; normalized pairwise distances are averaged.

Formula: mean pairwise distance among the three section profiles.
Evidence anchors: 0.25 = 0; 0.20 = 1.

Controlled comparison

Eight writing genres

Using different genres reduces the chance that one topic or writing format determines the leaderboard.

SEO article about coffee roasting
Academic essay on universal basic income
Research writing about AI detection
Personal writing about overcoming insecurity
Fiction about a Halloween party going wrong
Biography of Alan Turing
Copywriting email for a dating service
News report on New York's cost of living

Questions

About the LLM writing benchmark

Does a high score prove that a text was written by AI?

No. The score compares measurable stylistic tendencies. It is useful for controlled model comparison, but it is not proof of authorship for an isolated document.

Why is the human baseline included?

The human samples provide a practical reference under the same eight genres. They make the direction and scale of each model result easier to interpret.

How often does the leaderboard update?

The website reads the latest benchmark published from the Intellectualead app. Completed NLP backfills and new benchmark runs publish refreshed metrics automatically.