GLTR Reviews

GLTR Description

GLTR is designed to utilize the same models that generate counterfeit text as a means for detection. It incorporates the GPT-2 117M language model from OpenAI, which stands out as one of the most substantial models accessible to the public. By taking any given textual input, GLTR evaluates the predictions made by GPT-2 at each position in the text. The output showcases a ranking of all words recognized by the model, allowing us to determine how the actual following word ranks in comparison. Utilizing this positional data, a color-coded mask is applied to the text, reflecting the ranking position of each word. Words that rank among the most probable are shaded in green (for the top 10), yellow (for the top 100), red (for the top 1,000), while the remaining words appear in purple. Consequently, this method provides a clear visual representation of how probable each word is according to the model's predictions, ultimately enhancing our ability to identify potentially fake text. Additionally, this visual tool can help users quickly gauge the authenticity of a given passage.