What does ChatGPT Understand? A Challenge for the Strategy

Lisa Miracchi Titus, University of Denver

Thursday, October 5, 4 pm @ LSC 312

Can ChatGPT (and similar AI systems) understand the prompts they are given? Do they understand their own responses? I argue that, despite their impressive behavior, the way ChatGPT and similar models are trained make them unlikely candidates for semantic understanding. Unlike humans and many animals, who are causally sensitive to the meanings of the words we use, ChatGPT is plausibly just sensitive to statistical information about text co-occurrence. I conclude by discussing some implications for AI research and its ethical application.