Stylistic Signatures of LLMs and How to Detect Them

Date:


Abstract

In this talk I summarize what we know and understand about the stylistic tendencies of Large Language Models (LLMs) from a linguistic and statistical perspective. I also discuss both why it is unintuitive that LLMs have a style to begin with and some theories as to where those stylistic tendencies actually originate from. I conclude by discussing the implications this has on our ability to robustly identify AI-generated text.

Location

This talk was given on February 7th, 2025 at the University of Pennsylvania in Philadelphia, PA as part of the weekly Integrated Language Science and Technology (ILST) Seminar.

Slides