GLTR Reviews

GLTR Description

GLTR's goal is to use the same models that were used to generate fake text as a tool to detect it. GLTR has access the GPT-2 117M language modeling from OpenAI. This is one of the most widely available models. It can analyze any textual input to determine what GPT-2 would predict at each position. We can calculate how the following words rank because the output is a ranking all the words the model knows. This positional information is used to overlay a colored mask on the text that corresponds with the rank. The most likely words are highlighted in green (top 10) or yellow (top 100), while the rest are highlighted in purple. This gives us a visual indication of the likelihood that each word is under the model.