try it
two demos, five models
Both run on data from my papers on whether a language model leaves a fingerprint in what it writes. In the first, you guess which model wrote an essay. In the second, a paper’s classifier reads an essay, or your own text, in your browser.
[01]
Which model wrote this?
// langllm · ten essays · about two minutes- 1Read the essayA real answer from GPT, Claude, Gemini, Grok or DeepSeek to one of the twelve prompts in the LangLLM study.
- 2Name the modelPick one of the five. A blind guess is right 20% of the time.
- 3See what gave it awayWhat my classifier said and the features it went on, and what each of the five models guessed.
0/0you
0/0my classifier
0/0the five models
Loading…
Ten English essays, two per model, drawn at random. Data and code: LangLLM on GitHub ↗
[02]
Run the classifier
// compllm · 18 features · in your browser- 1Load an essay or paste textTen essays from the CompLLM study, two per model, or anything you write. Nothing leaves your browser.
- 2It measures 18 featuresThe paper’s own: length, sentence and paragraph shape, vocabulary, hedging, punctuation and readability.
- 3It names the closest modelA logistic regression fitted to the study’s 190 essays. With each essay’s prompt held out, it names the right model 87% of the time.
The paper’s 18 features and a logistic regression fitted to its 190 chat-app essays. Data and code: self-recognition-peer-baseline on GitHub ↗