Consider a common experience. You put a question to one of these systems — a large language model — something you had been stuck on. A problem at work, perhaps, or a matter in your own field that you had half-understood for years and never quite pinned down. And it answered. Not vaguely, but clearly and in order, with the very example you had needed and never found, the one that made the point clear. It had explained your own work to you better than the last several people you had asked.
Most people who have had that experience also have a second, immediate reaction: it does not actually understand any of this. It is a system that predicts the next word, trained on an enormous quantity of text, and there is no one inside it who knows what a single word means. That reaction is widely held, and it is meant seriously.
But there is a further response, and it is harder to set aside. If that is all the system is doing, how is it getting things right? If there is no understanding inside it, if it is only manipulating symbols it does not grasp, how did it find the exact example that was needed, and catch the point that others had missed?
Suppose you try to settle the matter — to say plainly what the machine is missing. The machine does not understand the word "gravity," you might say, because it has never dropped a glass or felt the ground. But you have never been to the centre of a star or watched a galaxy bend light, and you understand "gravity" perfectly well. Very well, you say — because for the machine the word is not connected to anything real; it is only one token among others. But you learned many of your own words the same way the machine did: from other words, from being told, without ever encountering the thing itself. Each time you try to state the difference — the difference between what you do with a word and what the machine does, the difference that makes your words mean something and its words mere noise — the test you reach for turns out to disqualify some of your own words too.
That is where this series begins. Not with the machine, but with a difficulty of our own: after a good answer, we find we cannot say clearly what we are nonetheless sure must be true.
This is Philosophy for Us — philosophy for everyone, no degree required.
This is the first of nine episodes, and they concern that difference we could not state. Not whether the machine is conscious — whether there is anyone inside it — which is a separate question we will deliberately set aside. The narrower and harder question underneath it: does the machine mean anything it says, or does it only produce the words?
Here is what these nine episodes do, and it is not to settle whether the machine merely parrots or genuinely understands. We will take the most obvious fact about you — that you know what your own words mean, that whatever else is uncertain, surely that much is secure — and offer you a sentence to complete: the machine does not understand its words because… You will find you cannot complete it cleanly. And in failing to, you will notice something you have carried all your life and never stated: a whole account of what it is for a word to mean something. I do not raise this from any settled position of my own; I cannot state the difference either. So tonight we look for the distinction we could not draw, and see why the blank where the answer should go is so much harder to fill than it appears.
Let us begin with the two reactions, because most people have both, and the difficulty is that both are reasonable.
The first says the system is a parrot — an impressive one, but a parrot. Here, roughly, is what it does; we do not need the engineering to feel the force of it. It has read an enormous quantity of text, more than any person could read in a thousand lifetimes, and it has become extremely good at a single task: given some words, predicting which word tends to come next. That is the whole of it. When it discusses gravity, it is not consulting an understanding of gravity, as a person would; there is no such understanding to consult. It is calculating which words are likely to follow which, and its calculations are very accurate. So when it sounds wise, what we are hearing is not wisdom but a recombination of the many wise things that people have written, assembled into a sentence. There is no one behind it who means any of it.
This is not a naïve view. People who know exactly how these systems are built hold some version of it, and they are not being cynical. It is genuinely strange, and worth holding onto, that something with no grasp of anything at all can produce so much that is correct.
But here is the second reaction, and it is the one that is harder to dismiss. If that is all it is doing, how is it getting things right?
"Predicting the next word" sounds so deflationary, so mechanical, that it is easy to overlook what doing it well would require. To reliably predict the next word in a real explanation of a difficult idea — to find the single example that makes it clear to one particular person, to catch a subtlety that several experts had missed — the system must be keeping track of something. Whatever we call it, it behaves consistently as though it is following the point. It corrects itself. It responds to what a person actually meant rather than what they literally typed. And at that point "it is only predicting the next word" stops sounding like a deflation and begins to sound like a puzzle: if that really is all it is doing, why does it keep producing exactly the result that, in a person, only understanding produces?
So there are two reactions. It is only a parrot. And: a parrot could not do this. Neither can be silenced, because each answers the other at its weakest point, and neither finishes the argument.
Now, here is what I will not do. I will not choose one of them for you tonight and offer the reassuring conclusion — that of course it is only a parrot, and you were right all along. Nor will I offer the unsettling one — that it secretly understands, and you were wrong to doubt it. Both are verdicts, and both, I think, are premature, for the same reason. Consider what happens when you try to defend either.
To defend the parrot view, you have to say what the machine is missing — what understanding is, that it lacks. To defend the second reaction, you have to say what it has — the same thing, now named. Either way, the whole question turns on a single word that neither reaction has paused over: understanding. Both use it as though its meaning were obvious, as though everyone knows what it is to understand a word and the only open question is whether the machine meets the standard. But no one has said where the standard lies, or what it is.
And that is the difficulty from a moment ago — the distinction we could not draw. The distinction supposed to run between your words, which mean things, and the machine's, which only sound as though they do. We are sure the distinction is there. What we cannot do — what neither reaction can do — is say where it falls.
So before we can take either reaction seriously, before we even let the parrot view make its best case in the next episode, we have to examine the thing both reactions rely on and neither has inspected. Not the machine — the standard. What it is to mean a word at all. And the quickest way to discover that one has no clear account of it is to notice that one has been using an account all along, quietly, and to state it aloud.
There is a version of this argument you have probably already had, aloud or in your own head, and I want to set it carefully aside — not because it is mistaken, but because it is a different question, easily mistaken for this one.
That version asks: is anyone in there? Is the machine conscious? Is there something it is like to be it — some awareness behind the words, someone experiencing the conversation from the inside? The answer most people reach, and I would reach it too, is: almost certainly not. There is no reason to think there is any inner life there at all. It is not conscious; no one is experiencing anything.
If you have followed this series from the start, you have considered this before; we spent an earlier stretch on exactly that question — on what test we were quietly using to decide whether a mind was present. So it may seem we have already been here.
We have not. And this took me some time to see. Whether anyone is inside, and whether the words mean anything, are two different questions, and settling the first does not settle the second.
Consider. Suppose we grant the whole point about inner life: there is no one inside, no awareness, no experience whatever. Does that establish that the machine's words mean nothing? Stop and check, because the answer is not obvious. A book has no inner life, and the sentences in it mean things. A stop sign is not conscious, and it means stop. A note left on the kitchen table goes on meaning "back by six" long after the person who wrote it has stopped thinking about it — it sits there, an object with no awareness at all, meaning something. So meaning cannot simply be a matter of inner awareness — of something being consciously present to a mind — because many things with no awareness behind them carry meaning nonetheless. Which is to say: we can be entirely certain that no one is inside the machine and still be entirely stuck on whether what it says means anything.
That is the reversal, and it is the reason for this series. The question of mind — whether it is conscious — we are setting down. Not because it does not matter, but because we can answer it "no" and remain exactly where we started on the question that actually troubles us after a good answer: not whether something is conscious within it, but whether the machine understands the words, or only produces them. We weighed whether there was a mind inside. Here is the harder question — what is our test for whether our own words mean anything?
And once the question is put that way, the matter looks quite different. For if meaning is not a matter of inner awareness — ours or anyone's — then what is it? Where does it reside? When your word "gravity" means gravity, what exactly is that? Is it a connection between the word and the actual thing in the world? Is it the way you use the word, the moves you can make with it? Is it something occurring in your mind as you say it — which we have just seen cannot be the whole story? We have no ready answer. And this is not a gap in anyone's education. No one is ever taught it. We learn to talk without once being told what it is we are doing when our talk means something. We simply do it, all day, and the doing feels entirely transparent.
So the distinction we could not draw — between your meaning and the machine's — does not wait on further facts about the machine. It does not depend on what the machine is; it depends on what meaning is. And that is a question about us, and about words, and about the world — one we have been answering all our lives without ever reading.
Let me show you. I am going to offer you a sentence, and all you have to do is complete it.
Here is the sentence. The machine does not understand its words because ___. That is all. Complete it. You believe the first half completely — it does not understand. So you should be able to say why. Fill in the blank.
Most people reach first for the same thing, so let us reach for it and examine it. The machine does not understand its words because they are not connected to anything real for it. When you say "rain," there is actual rain at the other end; you have stood in it, it is wet, the word is fixed to the world by a lifetime of contact. The machine has only ever had other words — words defined by other words, never once making contact with the thing itself. That is a good answer. It feels exactly right.
Now hold it against your own vocabulary, and it begins to come apart. You understand "gravity." Have you been to the centre of a star? You understand "the Cretaceous" — a period that ended sixty-six million years before there was anyone alive to see it. You understand "Cleopatra," "a quark," "the bottom of the ocean," "justice." How many of those are fixed to the world by your own contact? You never touched a quark. You were not present for Cleopatra. You acquired those words exactly as the machine acquired all of its words — from other words, from being told, from the company each word keeps with other words you were also told. So if "not connected to anything for it" is what disqualifies the machine, many of your own words fall on the same side as the machine's, and you must either grant that the machine understands after all, or concede that most of your own words fail the same test. Neither is comfortable.
Very well, you say — that is not quite what I meant. Try the second answer. The machine does not understand its words because it does not really grasp them; it only uses them. It manipulates the words correctly without any hold on what they mean. That feels right too: there is plainly a difference between knowing how to use a word and merely emitting it.
But now consider a question, and try to answer it rather than pass over it: in your own case, where is the grasping? When you understand "gravity," what exactly is the grasp — the thing beyond all the ways you can use the word, discuss it, reason with it, apply it? Point to it. Not the using — the grasping beneath the using. Try, and you will find it surprisingly hard to name anything there that is not simply more using. I am not saying there is nothing there. I am saying that the moment you tried to name what the machine lacks, you assumed there is a "grasping beneath" the competent use, separate from it — and that is not a fact you checked. It is a picture you are holding. We will return to it; for tonight, simply notice that you reached for it.
And perhaps you try a third answer. It does not understand because it has no idea what it is trying to say — there is no intention behind the words, no one meaning anything by them. That is a different answer again, and it may be a better one. But notice that it is a third account of what meaning is. The first said meaning is a connection to the world. The second said meaning is a grasp beneath your use. The third says meaning is a matter of intention, of trying to say something. These are not the same idea; they pull in different directions. On the first, many of your own words are in trouble; on the second, the very thing you are invoking is hard to locate even in yourself.
And that is exactly the thing I said we would find. You completed the blank. You had an answer — in fact you had three. And every one of them quietly assumed a whole theory of what it is for a word to mean something, a theory you never chose. You did not weigh it against the alternatives. You did not know it was one option among several. You reached for the first account of meaning that came to hand and used it to sort yourself onto one side and the machine onto the other. We all do this, all our lives. It goes unnoticed because it works — until a machine begins using words as fluently as we do, and we go to say why that does not count, and find we have assumed a theory we did not know we held.
So here is the one thing this series does, stated as plainly as I can state it. We are going to find the account of meaning you are already using and have never stated aloud — and then ask whether it is the right one, and whether it does what you have been trusting it to do, which is keep you on your side of the distinction. That is the task. Not to decide whether the machine understands, but to find out what you meant by "understand" before you let the word out of your mouth.
Let me state plainly what we have now that we did not an hour ago, because the gain is worth being clear about, even though nothing has been settled and I promised no reassurance.
You came in holding a dispute: is the machine a parrot, or does it understand? Mere prediction, or genuine understanding? And a quiet worry that you could not close it. You leave knowing that this was not yet the dispute to have — not because either side is wrong, but because both depended entirely on a word neither had defined: understand. You cannot say whether the machine understands until you can say what understanding is; and the moment you tried to say what it is, you found you had been carrying an answer all along and had never inspected it. That is the gain, and it is a genuine instrument, so let me hand it over plainly. You can now catch yourself in the act. You can notice the moment you settle on some account of meaning — it is not connected to the world, it does not really grasp anything, it does not mean to say anything — and stop and see that you selected one account out of several, without examining it. An hour ago that was invisible. Now you can watch it happen. That is not nothing. It is the difference between being certain and knowing why you are certain — and discovering that the two are not the same.
Here is where we stand. This is the first of nine episodes, and tonight was the foundation for the rest: not an answer, but a better question, and the one habit needed to ask it honestly. The questions become more difficult from here.
Next time, we stop declining the parrot view. We take it up and state it at full strength — not the lazy version, it is only autocomplete, but the serious one: the case, built by people who understand exactly how these systems work, that something trained on nothing but the form of language can never reach its meaning. It is a serious case. It is the strongest form of the first reaction you walked in with, and next time it has the floor, with no interruption from me. Only after it has made its best case do we ask whether it in fact holds.
Before then, try the exercise once on your own, because it becomes yours only when you run it with no one setting it up for you. Choose a word — one you are quite sure you understand, but one you learned entirely from other words and never from the thing itself. "The Cretaceous." "A quark." "Inflation." "Justice." Take whichever is yours. Now complete the sentence for that word: the machine does not understand it because ___. And whatever you reach for, do what tonight has taught. Ask which account of meaning you have just assumed. If you said "it is not connected to the real thing for the machine" — were you ever connected to a quark? If you said "it does not really grasp it" — find your grasp of it, the part that is not simply everything you can say and do with the word. Do not settle it. I will not settle it for you, because the answer is only yours if you find it. Simply watch yourself assume a theory, and watch which side of the distinction it places you on.
For here is the thought I will leave you with. You still cannot draw the distinction. You know it is there — you know your "gravity" means gravity and the machine's "gravity" is something you are not prepared to call meaning — and you cannot, tonight, say what the difference is without assuming an account of meaning you have not earned. And here is what has changed: you now know the reason you cannot draw it is not that you know too little about the machine. We could open it up and learn every detail of how it works, and the distinction still would not appear, because it was never about the machine. It was about a question you have answered all your life without reading it — what it is for a word to mean something. Perhaps, when we are done, the distinction holds, and you keep your side, and the difference is real and can at last be named. Or perhaps some of the distinctions you have been drawing run straight back through your own words. I genuinely do not yet know which, and anyone who claims to be certain has not checked their own answer either. That is the question the next eight episodes take up. What happened tonight was not a failure to settle it. It was a question you had never actually asked coming clearly into view for the first time — and once it is in view, it is not easily set aside again.
Thank you for listening. I will see you next time.