GLTR (Giant Language Model Test Room)

On the record

Audience: The paper describes its detection methods as simple and explainable to non-experts.arxiv.org · read 5 Oct 2026
Availability: The project paper describes GLTR as open-source and publicly deployed.arxiv.org · read 5 Oct 2026
Customization: Developers can add a custom model by implementing a backend API that inherits from AbstractLanguageChecker and registering it.github.com · read 5 Oct 2026
Detection methods: It applies baseline statistical methods to identify generation artifacts across common sampling schemes.arxiv.org · read 5 Oct 2026
License: The repository identifies the project license as Apache 2.github.com · read 5 Oct 2026
Models: The project’s server supports gpt-2-small, BERT, or a custom model.github.com · read 5 Oct 2026
Purpose: GLTR is a tool to help people assess whether text was generated by a language model.arxiv.org · read 5 Oct 2026
Self-hosting: The repository documents installing dependencies and running the server locally.github.com · read 5 Oct 2026
Study result: In a human-subjects study, GLTR’s annotation scheme raised fake-text detection from 54% to 72% without prior training.arxiv.org · read 5 Oct 2026
Web demo: The repository links to an online demo at gltr.io/dist/index.html.github.com · read 5 Oct 2026