Mr. Seán O'Sullivan:

In February 2023, I tried to train my own AI content detector because I found that existing detectors could be bypassed through obfuscation and paraphrasing and other adversarial approaches like that. I found that even after training a model on specific paraphrased AI data, it could be bypassed through instructing the model to mimic an author's style. Even as the model moves from GPT-3 to the more capable GPT-4, and to GPT-4o and all the other frontier models, not just from up in AI, this system was built by taking a step back and asking what are we trying to do when we detect AI-generated content. We are trying to attribute authorship, and that is what I explored in this project.