GLTR
Forensically detects auto-generated text.
About GLTR
GLTR is a cutting-edge tool developed by the MIT-IBM Watson AI lab and HarvardNLP to automatically detect generated text through forensic analysis. By assessing the likelihood that a language model has produced the text, GLTR can pinpoint artificially generated content.GLTR visually dissects the output of the GPT-2 117M language model from OpenAI, enabling it to rank each word based on the model's potential production. The tool highlights words in green, yellow, and red based on their probability of being generated by the model, with the remaining words in purple. This intuitive color scheme provides a clear visual representation of the text's authenticity.In addition, GLTR showcases three histograms that summarize information across the entire text. These histograms reveal the distribution of words in each category, the ratio between the top predicted word and the subsequent word, and the entropy distribution of predictions. By examining these histograms, users gain further insight into the likelihood of a text being computer-generated.GLTR is a valuable tool for identifying fake content such as reviews, comments, or news articles created by sophisticated language models. These models can produce text that is virtually indistinguishable from human writing, posing a challenge for non-expert readers.Access GLTR through a live demo and find the source code on Github. Researchers can delve deeper into the tool's capabilities by reading the nominated ACL 2019 demo track paper.