Ghostbuster

Ghostbuster is a model-agnostic detector designed to detect AI-generated text. It works by passing input text through a set of weaker language models and then performing a structured search on the resulting features across all the models. A classifier is then trained on the selected features to make a prediction about the text's origin.

The detector can be useful for detecting AI-generated content, but the tool is not foolproof and results may vary depending on a number of factors. For example, it may not work well for shorter texts (less than 100 words), texts from domains that are different from those in its training data (e.g., text messages), non-American and non-British English texts, texts written by non-native English speakers, and texts that have been edited or paraphrased by humans after being generated by AI.

Ghostbuster is also currently limited by the deprecation of some OpenAI endpoints, which means it may not work as well. So be careful using the tool in real-world situations, especially if you're using it to automatically punish writers for using AI-generated text.

The tool sends input data to the OpenAI API and stores it for internal testing purposes. But users should be aware that input privacy is not guaranteed. For more details and to try the tool, check out the Ghostbuster website.

Published on June 13, 2024