AI Tech Suite

GLTR

Click to visit website

About

GLTR (Giant Language Model Test Room) is designed to inspect the visual footprint of automatically generated texts. Developed by Hendrik Strobelt and Sebastian Gehrmann, with review by Alexander Rush, it's a collaboration between MIT-IBM Watson AI Lab and HarvardNLP. The tool leverages the capabilities of large language models, specifically utilizing OpenAI's GPT-2 model, to detect whether a text was generated by a machine rather than by a human. It analyzes given text by comparing the likelihood of each word's occurrence against the predictions of the language model, providing a visual representation of its findings. Words are color-coded based on their predictability—green for top 10 predicted words, yellow for top 100, and red and purple for less likely predictions. This forensic analysis aids in identifying potential misuse of AI-generated text for malicious purposes like fake news or reviews. Furthermore, GLTR is open-source and has been publicly deployed for any users interested in detecting machine-generated text.

Features

• interactive predictions preview

• histograms showing prediction distributions

• equipped with statistical detection methods

• color-coded visualization

• detects machine-generated text

FAQs

What is GLTR used for?

GLTR is used for detecting automatically generated text by analyzing its predictability and visual footprint.

How does GLTR work?

GLTR utilizes a language model to predict word likelihoods and overlays a colored mask on text to indicate how likely each word was generated by the model.