GLTR

About GLTR

GLTR, a collaboration of MIT-IBM Watson AI lab and HarvardNLP, is a tool introduced to inspect the visual footprint of automatically generated text. It enables a forensic analysis of how likely an automatic system has generated a text.

Background and Development

Developed by Hendrik Strobelt and Sebastian Gehrmann, GLTR was introduced as a tool to prevent the misuse of large language models for generating fake text. The tool assumes that computer-generated text fools humans by sticking to the most likely words at each position, a trick that fools humans. In contrast, natural writing selects unpredictable words that make sense to the domain.

Core Features and Capabilities

GLTR uses the GPT-2 117M language model from OpenAI for detection. It can use any textual input and analyze what GPT-2 would have predicted at each position. The output is a ranking of all of the words that the model knows, and this positional information is used to overlay a colored mask over the text that corresponds to the position in the ranking.

User Experience

Users can paste any text into the tool for analysis. The tool provides a visual indication of how likely each word was under the model. By hovering over a word in the display, a small box presents the top 5 predicted words, their associated probabilities, as well as the position of the following word.

Applications and Use Cases

GLTR can be used to detect fake reviews, comments, or news articles. It can also be used to analyze text from other automatic text generation systems.

Impact and Future Outlook

GLTR represents a visually forensic tool to detect text that was automatically generated from large language models. Despite its limitations, it can spark the development of similar ideas that work at greater scale.