Skip to content
Imtiaz Hossain

playground / nlp / stylometry

AI-Generated Text Detection

Twenty-eight statistical features decide human vs machine. Paste text or pick a sample and watch the classifier reason.

Full case study →Report (PDF) ↗
HC3 accuracy
97.40%
Random Forest
macro F1
0.9700
Random Forest
cross-gen
20-30%
on Bloomz, below chance
top feature
whitespace
MI 0.255
train corpus
HC3
/ 84,391 texts, 5 domains
features
28
/ lexical / syntactic / readability / distributional
selection
mRMR + RFE
/ each picks 15
compute
CPU-only
/ no GPU inference

interactive / live in your browser

or try

top discriminative features / mRMR

whitespaceautomatedvocabularyzipfhapaxsimpsonsyules

Whitespace ratio, readability, and vocabulary richness carry the most signal.

the pipeline

From raw data to a verifiable result

  1. 01 / dataset

    HC3 human vs ChatGPT

    The Human ChatGPT Comparison Corpus pairs human-expert and ChatGPT answers across open QA, finance, medicine, wiki/CS-AI, and Reddit ELI5. After filtering, 84,391 texts: 57,552 human, 26,839 AI.

  2. 02 / features

    28 stylometric features

    Every text is reduced to 28 model-agnostic surface features across four categories, capturing vocabulary richness, punctuation habits, readability, and distributional structure.

    8 lexical7 syntactic5 readability8 distributional
  3. 03 / selection

    mRMR and RFE agree

    Two independent methods, a filter (mRMR) and a wrapper (RFE), each select 15 features. Their agreement on whitespace ratio, vocabulary richness, and Zipf's coefficient is strong evidence those features genuinely matter.

  4. 04 / training

    Four classifiers, grid search

    SVM, AdaBoost, Decision Tree, and Random Forest, each tuned via 5-fold stratified GridSearchCV over a StandardScaler pipeline.

  5. 05 / inference

    Score a passage

    Below: paste text or choose a sample. The demo surfaces the nearest curated example, its verdict, and the feature signals that drove it.

  6. 06 / evaluation

    The generalization cliff

    In-domain the model is excellent. Move to a different generator (Bloomz) and it drops below chance. The feature fingerprints are generator-specific, not universal.

evaluation artifacts

next experiment

News Topic Classification