Detecting and Characterizing AI Use in Writing

Nowadays, lots of people are using AI to “help” write things. For instance, our recent study finds that approximately 9% of recently published American newspaper articles are produced either partially or entirely by large language models (LLMs). However, most of this AI use is undisclosed to readers, raising important concerns of transparency and public trust. In this talk, I will discuss the broad spectrum of AI-human writing, from minor edits for style/grammar to the generation of large blocks of content, in the context of what might be acceptable and unacceptable to a reader given lack of disclosure. I’ll then discuss EditLens, an automatic method that performs post-hoc AI detection: given a piece of text, can we tell to what extent AI was used to assist in its writing? I’ll then discuss an extension of EditLens that aims to separate AI ideation from AI writing, and qualitatively characterize the space of AI ideas compared to human ideas. Finally, I’ll describe ongoing work measuring the impact of AI-generated text during pre-training on LLM capabilities. We show that AI text is “worth less” than human text during training, which ultimately necessitates modifications to standard scaling laws.