Few words divide the field as much. In a 2022 survey of natural-language-processing researchers, cited by M. Mitchell & Krakauer (2023), 480 respondents split almost evenly (51% against 49%) on the claim that a model trained only on text could, given enough data and compute, understand language in some non-trivial sense.
The skeptics offer arguments of different kinds. For Searle (1980), formal symbol manipulation yields no understanding at all: the program's understanding is not partial but "zero" (1980, p. 419) (see The Chinese Room). For Bender & Koller (2020), to understand is to retrieve the communicative intent behind an expression, which cannot be learned from form alone (2020, p. 5187) (see form vs meaning and The Deep-Sea Octopus). M. Mitchell (2019) speaks of a "barrier of meaning": understanding a situation requires core knowledge of intuitive physics and psychology, mental models to simulate and predict, abstraction and analogy, and no AI system yet possesses it (2019, ch. 14).
Millière & Buckner (2024a) propose sidestepping the verbal dispute and speaking of semantic competence, distinguishing, after Marconi, its inferential aspect (defining, paraphrasing, drawing consequences) from its referential one (connecting words to objects and actions). The first seems within reach of LLMs; the second is more contested (2024a, sec. 3.3). They also warn against the "Redescription Fallacy": that a system can be described as "mere statistics" or "next-token prediction" does not show that it cannot implement a cognitive capacity, just as describing a piano as hammers striking strings does not show that it cannot produce harmony (2024a, sec. 3).
M. Mitchell & Krakauer (2023) suggest that the problem may be treating understanding as a single thing: there could be distinct modes of it, each with its own strengths and limits. The underlying question is what kind of fact understanding is: a behavioral capacity, an internal model with a certain structure, or something that requires intentionality and consciousness?