GLTR

What Exactly is GLTR?

GLTR, or “Catching Unicorns With GLTR,” is a tool developed by the MIT-IBM Watson AI Lab and HarvardNLP. Its purpose is to spot text generated automatically, specifically by OpenAI’s GPT-2 language model. It works like a highlighter for AI-written words, showing how likely each word in a text was generated by a computer using a color-coded system. The goal is to make it easier for non-experts to tell artificial text apart from human writing, supporting transparency and trust in the information we read.

The tool uses statistical methods based on word probabilities and applies a colored overlay to the text. Green means the word was among the top 10 most likely predictions, yellow for the top 100, red for the top 1,000, and purple means it was less likely to be predicted by the model.

GLTR also serves as an educational tool. It offers examples of both real and AI-generated texts, helping users understand how language models behave. It is freely available online and supports forensic analysis of text created by these models.

Who’s Behind GLTR?

GLTR was developed by Hendrik Strobelt and Sebastian Gehrmann in collaboration with the MIT-IBM Watson AI Lab and HarvardNLP. It is a forensic analysis tool designed to detect automatically generated text, providing a visual “footprint” that helps distinguish human-written text from AI-produced text.

What Can You Do With GLTR?

• Act as a forensic tool to determine whether a text was written by a human or generated by a language model.
• Provide a visual way to inspect AI outputs using statistical detection methods.
• Use OpenAI’s GPT-2 117M language model to compare predictions with the actual text, analyzing how likely each word was automatically generated.
• Display histograms that break down word categories by probability ratios and prediction entropies for forensic analysis.
• Serve as an educational resource with samples of both real and fake texts.
• Accept pasted text for analysis, visually showing each word’s probability of being model-generated via the color-coded overlay.
• Help non-specialists identify automatically generated text, improving transparency and reliability in language processing.
• Offer free online access with a live demo for analyzing your own text.
• Help combat misuse of language models for creating fake content by making computer-generated text easier to detect.