Is an LLM 'intelligent'? How AI differs from a chair or a table

People ask me some version of this question constantly, usually after a model does something that feels uncanny: “Is it actually intelligent, or is it just faking it?” It is a fair question, and the honest answer is that it depends entirely on what you mean by “intelligent.” So let me try to pin the word down, and then place a large language model somewhere sensible on the map — somewhere between a table and a person, which is a wider gap than it sounds.

Start with the chair

A chair is inert. It has no internal state that responds to the world. Push it, and physics happens; it does not processanything. A table is the same. These objects have structure and even purpose in the sense that we designed them for a job, but the “purpose” lives in our heads, not in the wood. There is no sense in which a table takes an input, transforms it according to learned patterns, and produces an output. It just sits there being a table.

This is the first thing an LLM clearly is not. A model is a function that maps input to output through hundreds of billions of learned parameters. Feed it text and it does something a table can never do: it computes a response that is contingent on what you said. That alone puts it in a different category. But being non-inert is a low bar — a thermostat clears it. The interesting question is how far past the thermostat an LLM actually gets.

What the model actually does

Underneath the conversational surface, a language model does one thing: it predicts the next token(roughly, a word fragment) given all the tokens so far. During training on an enormous body of text, it adjusts its parameters to make the true continuation more probable. After training, generating a sentence means repeatedly asking “given everything up to here, what token is most likely next?”, sampling one, appending it, and asking again.

context = "The capital of France is"

# the model outputs a probability over the whole vocabulary
p(next | context) = { " Paris": 0.91, " a": 0.02, " located": 0.01, ... }

# sample a token, append it, repeat
context += " Paris"

That is genuinely all that is happening at inference time. What makes it feel like more is the sheer richness of the distribution it learned. “What word comes next” sounds trivial until you realise that to predict the next token well across billions of examples, the model had to encode a staggering amount of structure: grammar, facts, associations, reasoning patterns, the shape of an argument, the conventions of code. Predicting text at that level is not a parlour trick — it requires internal representations that behave, in many useful ways, like knowledge.

The Chinese Room

This is where philosopher John Searle’s Chinese Room argument earns its keep. Imagine a person who speaks no Chinese sealed in a room with an enormous rulebook. Chinese characters come in under the door; the person looks up rules, copies out other characters, and passes them back. To a Chinese speaker outside, the room answers fluently. Yet the person inside understands nothing — they are shuffling symbols by rule. Searle’s claim: the room manipulates syntax without ever touching meaning, so passing the test is not the same as understanding.

It is a sharp intuition, and it maps cleanly onto an LLM: symbols in, symbols out, no obvious place where meaning “lives.” But the argument has a famous reply worth holding alongside it. Maybe no single component understands, and yet the whole system— rulebook, process, and the structure encoded in it — does something we would be stubborn to call anything other than a kind of understanding. When a model correctly answers a question no one wrote down, generalising from patterns rather than looking up an answer, “pure symbol shuffling” starts to feel like an incomplete description.

Powerful, but not a mind

Here is where I try to be precise instead of poetic. A model can be extraordinarily capable and still lack the things we usually bundle into “intelligence” when we say it about a person:

• No persistent self.Between calls it has no continuous experience. Each request reconstructs context from scratch; there is no “it” waiting between your messages.
• No goals of its own. It does not want anything. Any apparent goal is the shape of the training data plus your prompt, not desire.
• No grounding in the world. It learned from text about reality, not reality. It has never touched a table.
• Confident when wrong. Because it optimises for plausible continuations, a fluent falsehood and a fluent truth look identical from the inside.

So “sentient” is the wrong word. There is no evidence of experience, only of competence. But “just autocomplete” is equally lazy — it undersells a system that has internalised more structure about language and the world than any human could enumerate.

Where the useful line is for builders

When I design systems, I skip the metaphysics and treat the model as what it demonstrably is: a very capable, ungrounded pattern engine. That framing predicts its behaviour better than either “it’s basically a person” or “it’s a fancy lookup table.” It tells me to give it the context it lacks (retrieval, tools, real-time data), to verify its outputs rather than trust its confidence, and to design for the fact that it has no memory or stake between turns.

Is an LLM intelligent? It is unambiguously more than a chair, and unambiguously less than the human reading this. The most useful posture is to stop trying to award or deny it the label and instead ask the only question that changes what you build: what does it reliably do, and what does it reliably not? Answer that, and the philosophy takes care of itself.