Does ChatGPT have semantic understanding? A problem with the statistics-of-occurrence strategy

Abstract

Over the last decade, AI models of language and word meaning have been dominated by what we might call a statistics-of-occurrence, strategy: these models are deep neural net structures that have been trained on a large amount of unlabeled text with the aim of producing a model that exploits statistical information about word and phrase co-occurrence in order to generate behavior that is similar to what a human might produce, or representations that can be probed to exhibit behavior similar to what a human might produce (meaning-semblant behavior). Examples of what we can call Statistics-of-Occurrence Models (SOMs) include: Word2Vec (CBOW and Skip-Gram), BERT, GPT-3, and, most recently, ChatGPT. Increasingly, there have been suggestions that such systems have semantic understanding, or at least a proto-version of it. This paper argues against such claims. I argue that a necessary condition for a system to possess semantic understanding is that it function in ways that are causally explainable by appeal to its semantic properties. I then argue that SOMs do not plausibly satisfy this Functioning Criterion. Rather, the best explanation of their meaning-semblant behavior is what I call the Statistical Hypothesis: SOMs do not themselves function to represent or produce meaningful text; they just reflect the semantic information that exists in the aggregate given strong correlations between word placement and meaningful use. I consider and rebut three main responses to the claim that SOMs fail to meet the Functioning Criterion. The result, I hope, is increased clarity about why and how one should make claims about AI systems having semantic understanding.