Links

SurfacePrint

Typographic Fingerprints of LLMs

Wentao Zhang, Sajad Ebrahimi, Nima Jamali, Bardia Shirsalimian,
Kelly McConvey, Jalehsadat Mahdavimoghaddam, Maksym Taranukhin,
Maura R. Grossman, Vered Shwartz, Ebrahim Bagheri, Yuntian Deng

Same word. Different apostrophe.

don't
U+0027
UTF-8: 27
don’t
U+2019
UTF-8: E2 80 99

Same words, different formatting

Plain text
A simple answer

It doesn't change.
Markdown
**A simple answer**

It doesn’t change.

Count the formatting.

’ ’ ’ ' '
3 / 5

60% curly

SurfacePrint

97 measurements of a text’s formatting

A random forest learns formatting patterns.

Formatting features feed a random forest, producing a human-to-language-model score Raw text becomes 97 formatting features. An ensemble of 300 decision trees combines those features into a continuous score, from more human-like to more language-model-like. The trees are schematic illustrations, not specific learned rules. Raw text don’t **bold** 97 features ··· Random forest 300 trees Score Human LM

Detect AI-generated text

TRAIN

Human text
+ 17 AI models

TEST

New texts

Evaluation

Human text<Generated text

Detector score

0.5Chance
AUROC
1.0Perfect

Formatting separates these benchmark texts.

0.983

Mean AUROC

Held-out texts · 17 known model sources

Even whitespace carries a fingerprint

Raw text
First paragraph.


Second paragraph.
0.983 AUROC
Whitespace cleaned
First paragraph.

Second paragraph.
0.889 AUROC

Instruction-tuned model generations are often easier to detect

Whitespace cleaned

Base modelAUROC 0.70.80.91.0 Llama 20.810Llama 3.10.744OLMo 30.709Qwen 3.50.840Gemma 40.754 Instruction-tuned/ 0.804/ 0.880/ 0.958/ 0.978/ 0.963

Different model families have different fingerprints

Train: human + Claude Sonnet 4 · cleaned text

0.00.250.5Chance0.751.0Claude Sonnet 4.6 (same family)0.969OPT-IML 1.3B (different family)0.301

Learn a new model family’s fingerprint

Collect
examplesHuman + new AI family
Retrain
SurfacePrintLearn its fingerprint
Detect
AI-generated textUse the updated detector

AI writing styles leave
revealing fingerprints.