Catching Unicorns with GLTR

GLTR, or the Giant Language Model Test Room, is an innovative forensic tool designed to help detect whether a text has been written by a human or generated by a language model. Developed through a collaboration between the MIT-IBM Watson AI lab and HarvardNLP, GLTR provides a visual footprint of language model outputs. Using the demo, users can input text and GLTR will analyze the likelihood that each word was automatically generated using a color-coded overlay to indicate probabilities. With access to GPT-2, GLTR utilizes statistical detection to help non-experts identify artificial text, fostering transparency and reliability in language processing.

Key Features

• Statistical Detection: GLTR uses baseline statistical methods to help identify generated text based on word probability rankings.

• Visual Footprint Analysis: A color-coded overlay visually depicts the likelihood that each word in a text was generated by a model.

• Access to GPT-2 117M: The tool uses the GPT-2 117M language model from OpenAI to check predictions against actual text.

• Histograms for Aggregate Data: GLTR presents histograms showing the distribution of word categories probability ratios and prediction entropies.

• Educational Tool: Provides insightful samples of both real and fake texts making it an educational resource for understanding language model behaviors.

FAQ

What is GLTR?

GLTR stands for Giant Language Model Test Room and is a tool for forensic analysis to detect if text is automatically generated or written by a human.

How does GLTR work?

GLTR works by analyzing text input and checking against predictions made by the language model GPT-2, then applies a color mask to indicate how likely it is that each word was generated by a computer.

What do the colors in GLTR represent?

Green indicates top 10 probability, yellow top 100, red top 1,000, and purple represents less likely predictions.

Can I try GLTR with my own text?

Yes, you can try GLTR by using the live demo available on the website and input your own text for analysis.

Is GLTR publicly accessible?

GLTR is available for public use and can be accessed online. It is also open-source with the source code available on GitHub.