2026 Summer Archive 3 — Can an LLM’s words refer to things in the real world?

Large language models learn from vast amounts of text and produce strikingly fluent responses. But if they process only patterns in language, how can their words refer to actual people, objects, or events?

Bender and Koller offer a skeptical answer. They argue that learning linguistic form alone cannot establish what those forms mean or what communicative intentions they express. Accurate text prediction is not, by itself, understanding.[1]

Mandelkern and Linzen challenge this argument in “Do Language Models’ Words Refer?” Drawing on linguistic externalism, they argue that training text is not a collection of symbols cut off from the world.[2]

Suppose someone says, incorrectly, that Peano proved the incompleteness of arithmetic. Gödel proved the incompleteness theorems, but the speaker’s use of “Peano” still refers to Peano. Reference is not determined solely by what an individual knows or believes.

On the externalist picture associated with Kripke and Putnam, reference also depends on the causal and historical chains through which words are used and passed on within a linguistic community.[3][4] Most people cannot reliably distinguish an elm from a beech, yet “elm” and “beech” still refer to different kinds of trees. Speakers inherit those connections from their linguistic community and its experts.

Mandelkern and Linzen apply this idea to LLMs. Training data was produced by people talking about real individuals, objects, and events. An LLM may not encounter those things directly, but it may still be indirectly connected to them through the history of the language in its data.[2]

It is therefore too quick to argue that an LLM’s words cannot refer to anything in the world simply because the model learns only from text. Text is not necessarily detached from reality. It records how people have classified, described, and referred to things.

For businesses, this is one reason to treat training data as more than raw strings. A dataset carries traces of the environment in which it was produced, the people who produced it, and the linguistic practices they followed. Its origin and context may help shape what a model’s outputs are about.

Still, reference is not the same as understanding. Whether an output refers, whether the model intends to refer, and whether it understands a concept are separate questions. Mandelkern and Linzen do not claim to have proved that LLMs understand meaning—or even decisively that their words refer.[2]

Their conclusion is narrower but important. We cannot infer that an LLM’s language is disconnected from the world simply because the model does not experience the world directly.

References

[1] Bender, E. M., & Koller, A. (2020). “Climbing towards NLU.” ACL 2020, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463

[2] Mandelkern, M., & Linzen, T. (2024). “Do Language Models’ Words Refer?” Computational Linguistics, 50(3), 1191–1200. https://doi.org/10.1162/coli_a_00522

[3] Kripke, S. A. (1980). Naming and Necessity.

[4] Putnam, H. (1975). “The Meaning of ‘Meaning’.”