Do Large Language Models Understand Language?

Ask whether large language models understand language and you will get confident answers in both directions. The confidence is the problem. The useful version of this question starts by admitting that “understand” is doing a lot of unexamined work.

This piece is part of a larger argument about how AI is changing human language. Here I want to slow down on the single hardest sub-question.

What a language model actually does

A large language model is trained to predict the next piece of text, over and over, across an enormous amount of human writing. In doing so, it builds an internal representation of how language fits together: which words follow which, how arguments are structured, what a question expects as an answer.

The surprising empirical fact, established by the transformer architecture and the scaling that followed, is that this single objective produces systems that can summarise, translate, reason through problems, and hold a coherent conversation. Nobody fully predicted that next-word prediction alone would get this far.

The case that they do not understand

The strongest skeptical argument is simple and hard to dismiss. A model has:

• No body. It has never touched, seen, or wanted anything. Its words point at a world it has no contact with.
• No goals. It is not trying to communicate. It is completing a pattern.
• No stake. It is indifferent to whether what it says is true.

On this view, the model is a remarkably good mimic of the surface of language with nothing underneath. The well-known phrase from one influential critique is that these systems are “stochastic parrots”: they recombine forms they have seen without grasping meaning.

The case that the question is wrong

The counterargument does not claim models are conscious. It claims the test is unfair.

If a system reliably uses language correctly across novel situations, at some point insisting it “doesn’t really understand” starts to look like moving the goalposts. We do not demand that a person prove inner experience before we credit them with understanding; we infer it from behaviour. Applied consistently, the same standard makes the model’s understanding at least an open question.

There is also a humbling possibility hiding here. If next-word prediction can produce this much competence, maybe a larger fraction of human language use is also pattern completion than we like to think. The model may be revealing something about us.

Why this is really a question about language

Notice what both sides assume. The skeptic says understanding requires grounding in a world. The defender says understanding is demonstrated through use. Underneath, they disagree about what language fundamentally is: a window onto private meaning, or a public system of coordination that runs largely on structure.

My own view, which I develop in Building Your First Brain, is that the truth is uncomfortable for both camps. Language carries far more structure independently of any single mind than the romantic view allows, which is why a model can wield it. But meaning still bottoms out in minds with stakes, which is why the model’s fluency feels hollow when you press on it.

That tension is exactly what becomes urgent when communication starts moving below the level of words, as it does in brain-computer interfaces.

A practical takeaway

You do not need to resolve the philosophy to use these systems well. The working rule that falls out of all this is: trust the structure, verify the substance. A model is reliable about the shape of language and unreliable about the truth of the world, because only one of those was in its training objective. Treat it accordingly.

For the long-form version of this argument, and where it leads once AI becomes a layer we think through rather than a tool we use, the book is free for the first 1,000 readers.

Further reading

• “Attention Is All You Need”, the architecture behind modern LLMs.
• “On the Dangers of Stochastic Parrots”, the influential case for caution about attributing understanding.