Jack Flintoft

The objection that many have against large language models being “intelligent” is that they do not understand the words which they are using (basically, that LLMs manipulate symbols without knowing what they ACTUALLY mean). 

I think that this charge only lands if we assume that human understanding is something more than pattern recognition. I’m not convinced it is. I’m rather willing to acknowledge that AI “understands” language in the same way we do, because human language itself seems to be a process of complex pattern-matching and association (refined by experience and feedback). Reading Quine’s philosophy for my class this quarter made this clear.

To frame this a bit better, many in the camp of negating LLMs ‘understanding’ of language seem to argue that humans (unlike AI) know fixed and deeply grounded meanings of words. On this view, we first internalize analytic truths, like: “I know that if someone is a bachelor, they are unmarried”. This analytic truth would have a 100% weighting. We then seem to go out and apply this internal knowledge to the world (“Tom is a bachelor, therefore he is unmarried”). The objection then, is that LLM’s skip the first step. They know how to ascribe “bachelor” to Tom without understanding what “bachelor” means.

Quine (a philosopher who wrote in the 1950’s) wanted to show that the supposed distinction between analytic truths (true by meaning alone) and synthetic truths (true by how the world is) simply doesn’t exist. Even our definitions — “bachelor,” “unmarried man” — are learned empirically, by observing how words are used. The meanings of words are socially constructed, not fixed.  What seemed true fifty years ago — that “bachelor” = “unmarried man” — has changed. Today, many unmarried men in lifelong relationships wouldn’t be called “bachelors”. Therefore, the previous weighting of 100% seems wrong …

The point here is that (much like a LLM) we as humans update meanings of words through shared experience in social environments, and never actually have “100% beliefs”.  

Quine’s thought experiment makes this a bit clearer. Say you, a linguist, trekked out to a new unknown land. A local is showing you around and “teaching” you the local language.

The local points to what you assume is a rabbit (for it looks like one) and says: “gavagai”. As a linguist, you take this to mean “rabbit”, but it could just as well mean “deconstructed rabbit parts” or “hopping” or “fluffy tail” or even “food that moves.” No amount of pointing and explaining will give cross us over the supposed threshold into rabbit becoming an analytic truth. The point is: we may never know what “gavagai” means. No amount of observation can tell us. 

Most of the time, we don’t notice this indeterminacy because we share enough of that web to communicate. But Quine’s point is that there is no final fact of the matter about what any word really means, only the patterns we converge on.

Coming back to the AI “understanding” point, when an LLM predicts the next word, it’s doing what we do: operating inside a web of associations, feedback, and use. It doesn’t need access to some “deeper” layer of objective meaning — because there isn’t one (as Quine shows). Human understanding itself seems to be a kind of probabilistic fluency, a skill in navigating and updating a network of relations between words, experiences, and expectations.

There appears, therefore, to be no ethereal realm where “true” meaning lives, only the endlessly shifting correlations between words, contexts, and experiences. If that’s parroting, then all of us are parrots — just highly sophisticated ones, trained by the contingencies of a human world.

---

Update 8/Nov/25: After some thinking, I want to clarify 2 things from this post after some interesting conversations I’ve had with people since:

(1) Someone objected to this article’s view by saying that LLMs don’t inherently understand the objects / things which they talk about. A thought experiment makes this argument a bit clearer:

Imagine a gaseous alien planet. Nothing really exists there other than gas. The aliens are made of gas, live in gas, and interact only with gas. It’s basically a world without objects. Now, imagine these aliens somehow come to “understand” the human language we use on Earth. One of them might say: “Chairs are wooden things” without ever experiencing wood or a chair. Because it seems, as humans, we seem to come to understandings of things through more than just language (through images, interactions, use). There’s a sense in which we embody the meanings of words.

The argument, then, is that these aliens (like AI) lack that embodied access. They may know how to say the right thing about a chair, but do they really understand what one is in this gaseous planet without chair objects?

But I think this kind of argument misunderstands how modern multi-modal models actually work (and where they are evolving towards). Even though an LLM doesn’t have a lived world to move through yet (it can’t see a chair in the way we can), concepts like ‘chair’ still live in a dense vector space — surrounded by images of chairs, videos of people sitting, transcripts, voice prompts, and more “embodied” stuff which isn’t just language. LLMs today are multi-modal and deeply entangled. I don’t think that’s a million miles away from how human understanding works either…

(2) Importantly, I actually think the mark and definition of intelligence has changed over the last few decades. By analogy, AI as a concept has morphed over the past 30 years (we used to call things like speech-to-text dictation software and Google Maps algorithms “AI”. Today, we just call them software. Once a system becomes understood, repeatable, and widespread, we quietly strip it of the “AI” label). Konstantine from Sequoia has an interesting article about this and more here

We seem to do the same thing with intelligence. As machines learn new capabilities (language / reasoning / creativity) we move the goalposts. We say, “Well, that’s not real intelligence.” But what we’re really doing is carving out a shrinking space that preserves something uniquely human.

Turing’s original behavioral test tried to bypass all of this. He wanted to judge intelligence by performance alone. But over time, it’s been revised, critiqued, and reframed by philosophers like Searle and Block, who shifted the conversation toward something more internal and less behaviourist … something that starts to look a lot more like consciousness. And that’s what I think we are being defensive about, not intelligence as such, but the feeling of having a mind (an important distinction…)

What I wanted to show with Quine and these thoughts is that if meaning / understanding itself is grounded in use then maybe there’s no deeper essence to access. I wanted to show somehow that AI is more intelligent (and closely resembling human intelligence [not consciousness]) than we normally like to think.