For four weeks now we have been trying to draw a distinction. Your words mean things; the machine's words only sound as though they do. And for four weeks, every place you have tried to draw that distinction, the test has failed in the same way. Say the machine's words are not connected to the world — and it turns out that half of your own words are not connected to the world either. Say it does not really grasp what it says — and when you look for your own grasp, the part that is something more than competent use, you find nothing there to point to. Every test you have raised to keep the machine out has, on inspection, put you on the same side as the machine.
So tonight I owe you the other half. It is not true — I said this last week, and I meant it — that every such test applies equally to you. There are tests that do not. There are serious philosophers, holding serious theories, who say the opposite: that there is something here, something you have and the system on your screen does not, and that no amount of fluent talk will ever supply it. Tonight those arguments get stated in full. And I want to be plain about one thing before they start, because it matters. It is going to feel like relief — four weeks of every distinction collapsing, and at last someone offers one that holds. Treat that relief as provisional. Because behind those arguments there is a second set, just as serious, who say that the question you have been asking all along — does it understand, yes or no — was the wrong question from the start.
This is Philosophy for Us — philosophy for everyone, no degree required.
This is the fifth of nine episodes. Last time, Ludwig Wittgenstein sent us looking behind our own use of a word for the understanding underneath it, and we came back empty-handed — in our own case as much as the machine's. Tonight, for the first time, the tests that do not do that: the ones that name a difference between you and the machine and make it hold. And then, before we get comfortable, the arguments that say the difference was never the thing we should have been looking for.
Let me begin with a frog, because the clearest version of tonight's first argument begins with one, and it is worth seeing where the idea comes from.
A frog sits still. A small dark thing moves across its field of vision, and the frog's tongue comes out and takes it. Inside the frog, when that dark thing moved, a particular signal fired — a state in its visual system that activates when, and roughly only when, something small and dark moves nearby. Here is the question that produced a whole theory of meaning. That signal — what does it mean? What is it about? The natural answer is that it means fly: it is the frog's detector for flies. But you could just as well say it means small dark moving thing, which is not the same, because a frog will snap at a pellet dropped past it as well. So which is it? What makes the signal about flies rather than about small dark dots? The answer this theory gives is the one the whole first half of tonight rests on. The signal means fly because catching flies is the job it was selected to do. Frogs that snapped at flies ate and bred; the detector is present in this frog because, over a long history, it tracked flies well enough to keep frogs alive. Its meaning is not fixed by whatever it happens to fire at on a given afternoon. It is fixed by what it is for — by what, across the history that produced it, it was built to track.
This is a serious position. The philosopher who developed it most fully is Ruth Millikan, in her 1984 book Language, Thought, and Other Biological Categories; Fred Dretske worked out a close version around the same time. The name for it is teleosemantics — from telos, meaning purpose or end, the job a thing is for. The core claim is simple: a sign means what it was selected to track. Meaning comes from function — from a thing's having been built, by some history, to do a job. The frog's state, a human word, a gene, an animal's warning cry: each means what it does because of what it was shaped to track, for some creature, across some history.
Now apply that to the two things in front of us, and notice the difference it makes — because for the first time in four weeks, the same objection does not apply equally to you.
You are a creature with a history. Not only your own life, though that too — you learned the word water as a small animal that got thirsty and was handed a cup, that splashed in it, that needed it and got it. Behind your life there is a longer history: you are the kind of thing you are because, for an immense stretch of time, the creatures you descend from had to track the world to survive — find the water, read the weather, tell the predator from the shadow. Your capacity to mean anything at all rests on that. Your words, on this account, are the recent, verbal outgrowth of a system that was built, over a very long time, to be about the world, because being about the world is what kept your line alive.
The system on your screen has none of that. It has a corpus — an enormous record of what people have said. That record is, to be fair, full of the world at second hand, the way a library is full of the world. But the system itself was never selected to track anything. It has no history of being built, by survival or need or any biological function, to track the things it talks about. It was trained, yes — trained to predict text; but that task was to get the next word right, never to be about water, or weather, or a predator in the grass. That is the only thing the history behind it ever shaped it to do. So on this account, when it says water, there is nothing its word was ever for — nothing it was built to track — and therefore, however perfectly the word patterns, there is nothing there for it to mean about. Your word came out of a history of needing the world. Its word came out of a history of needing the next word to be right.
Notice what has happened, because it is new: a test whose result does not apply equally to you. There does seem to be a difference here, and it falls the way you always wanted it to. You have an animal's past; the machine has a training run. You were built — by evolution, by a childhood, by a body in a world — to track things; it was built to continue text. So here, at last, is a place where the distinction holds. But everything turns on one word: it looks as though it holds. It divides you from the machine only if a history of being built to track the world is what meaning actually is. The relief and the condition arrive together — take them together.
I am not going to take the difference away, because it is real. But it comes with a cost, and the cost is the reason we are not finished.
You have just said that meaning comes from a history of being selected to track the world. Very well — but notice what you did. You chose that as what meaning is. There were other options. Last week the answer was use — meaning is how a word works in a practice, no history required, only competent going-on. The week before it was company — meaning is the pattern of a word among other words, which is the one thing the machine has in full. Tonight you reached past both and took function: meaning is what a thing was built to track. Each of the three is a serious theory. Each draws the distinction in a different place. Use and company both admit the machine, or nearly. Function keeps it out. And you reached for the one that keeps it out — on the very night you most wanted a difference that holds — without yet saying one word about why function, rather than use or company, is what meaning actually is.
That is the difficulty, and I want it stated clearly before we go on. The difference teleosemantics names is real. But it settles the matter only if teleosemantics is right — only if built to track is genuinely what meaning is, and not merely one of three live accounts that happen to disagree about exactly the case in front of us. That has not been established. No one in this argument has established it; that is why it is still an argument. So hold the relief and the cost together: there is a difference, and you cannot yet say it is the difference. Keep that, because the next argument names another one — and it is harder to set aside.
The second argument begins with a thought you can run in your own head right now, and it is the one that gave this whole problem its name.
Imagine you are handed a dictionary of a language you do not know a word of — Finnish, say, and you have no Finnish. You look up a word. The definition is, of course, in Finnish: more words you do not know. So you look those up. More Finnish. You can go on forever. Every word sends you to other words, and none of them ever sends you out of the book, to the thing it is about. You could memorise the entire dictionary, every definition perfectly, and you would know exactly how all the words relate to one another and still not know what a single one of them means — because you never once got out of the book. You never got from the words to the world.
That is the picture, and the philosopher who sharpened it into an argument is Stevan Harnad, in a 1990 paper; he called it the symbol grounding problem. The claim is this. Symbols defined only by other symbols are ungrounded — the closed dictionary, a web of words pointing at one another with no exit. And meaning, Harnad argues, cannot be only that. For words to mean anything, at least some of them must be grounded — tied directly to the things they pick out, not through further words but through a body's contact with the world. You know what red means, in the end, not because of a definition but because you have seen red — a perceptual system met the world and sorted it. You know what heavy means because you have lifted things. The grounding is the body discriminating, handling, coping. Some thinkers in this tradition — the enactivists — go further: meaning simply is a living body's coping with its surroundings, and you cannot have the meaning without the coping. Either way, what the words finally rest on is a body in a world.
Where the machine falls is now obvious. It is the person who memorised the whole Finnish dictionary. It holds every relation between every word — more than any human could — and, on this account, no exit from the book. It was trained on text: symbols defined by other symbols, the company a word keeps, all the way down. It has never seen red. It has never lifted anything. Its word heavy is tied to ten thousand sentences about heaviness and to not one thing it ever strained to pick up. So the symbols are ungrounded — the closed dictionary. However fluently they pattern, on this account nothing ties them to the world, because the one thing that could tie them — a body's contact with what the words are about — it does not have, and you do.
I have to be careful here, because this resembles something we heard a few weeks ago and it is not the same. This is not Searle's locked room from early in the series. Searle's point was about syntax — the shuffling of symbols by their shapes. And the well-known reply to Searle was: put the system in a robot, give it eyes and hands, and then perhaps the symbols acquire meaning. Searle only ever granted that a body might help. This argument says something stronger and cleaner: grounding in a body is not something that might help, it is the whole of what meaning requires at the bottom — without it you have the dictionary and nothing else. That makes it a sharper version of the difference than any we have had: not it is only shuffling shapes, but it has all the words and none of the world, because it has no body with which to meet the world, and you do.
That is the second difference that holds. And it comes with the same cost as the first — I will state it quickly, since you can see it coming. The difference is real if meaning requires a grounded body. But that is a claim, not a given. The use-theorist from last week would object: a body? I taught you that understanding is competent going-on, and there are plenty of your words — justice, the Cretaceous, inflation — that you handle perfectly and never grounded in any body's contact at all. So meaning requires a body is one more theory you would have to establish over the others. Treat it the same way: a real difference, and a cost not yet paid. And keep one further point in reserve, because we return to it shortly: this whole argument leaned on the phrase text-only. It assumed the machine only ever encountered symbols. Treat that assumption as provisional too.
Then there is a third argument, the one people reach for most often without knowing its name. It concerns intention.
The thought is this. When you say something to me — really say it, mean it — you are not merely producing a well-formed string of words. You are trying to do something to me. You want me to believe something, or feel something, or pass the salt. And there is a second layer that is easy to miss: you want me to recognise that you are trying to do it. You want me to take your intention as an intention directed at me. The philosopher Paul Grice, in a 1957 paper titled "Meaning," identified this and built an account of meaning out of it. To mean something, he argued, just is this: to say it intending to produce some response in a hearer, and intending that the hearer recognise that very intention. Meaning, on this account, is not in the marks or the sounds or the statistics. It is in one mind's directed reaching toward another — I want you to take this, and I want you to see that I want you to.
Apply that to the machine. When it produces a sentence, is there anything it is trying to get you to believe? Is there an intention directed at you that it wants you to recognise? On this account, no. There is nothing in the system doing the intending. It generates the string; you supply the sense that someone meant it. You read an intention into it — an intention that is not there. The fluency is real; the intention is yours. So it is not meaning; it is being read as though it meant.
You may be thinking: did we not already hear this? A few weeks ago, the critics who said the model only assembles form and that we project the understanding onto it. Yes — and here is the difference, which is worth a moment. That earlier claim was a charge: the model has no access to what it is talking about, so your sense that it means things is something you add. Whether or not it was true, it was an accusation. Grice supplies the theory beneath the accusation — the account of what meaning is that would make the accusation stick. The critics said it is not really communicating. Grice tells you what communicating is, such that the machine is not doing it: the reaching, the intending-to-be-recognised, and there is nothing inside to do the reaching.
And this is the part that gives it force: apply Grice to your own case and the objection does not follow. This is new. Every test for three weeks has applied equally to you. Not this one. You plainly do reach for other people. You plainly do say things wanting to be understood, and wanting the other person to catch that you want it. Whatever else is uncertain, that you do, all day. So here, cleanly, is something you have and the machine lacks, and the same test cannot be raised against you. It holds.
Three arguments now, and three differences that appear to hold. The machine was never built to track anything. It has no body with which to meet the world. There is no one within it reaching for you. After a month of every distinction collapsing, that is genuinely something. So let me state plainly what you are holding, before the next arguments arrive and tell us what is wrong with the question that got us here. You are holding three real candidates. Each names something the machine lacks and you have. And each is also a theory — function, body, intention — that holds the distinction only if it is the right account of what meaning is; you have established none of them, and three serious accounts cannot all be the one true thing meaning is. So what you have is real and unfinished at once: there is something here, and you still cannot say which of the three, if any, settles it.
So you are holding three differences. Let me bring in the arguments that dispute none of them — and still take the verdict away. They do it by going after what all three differences share: the shape of the question. Does it understand — yes or no. Watch what happens to that question when you look harder at it.
Begin with the easiest weak point, which is in the second difference, the one about the body. That whole argument rested on two words: text-only. The machine has only the dictionary, never the world, because all it ever encountered was symbols. I told you to treat those two words as provisional. Here is why. It is no longer true, or not straightforwardly. The systems people actually use now do not only read text. They take in images. They take in sound. They take in video. They run tools, they call other programs, they operate robot arms in laboratories and close a loop between a command, a camera, a gripper, and the world changing as a result of what they did. Researchers who work on grounding — tying symbols to things through a machine's own sensors and actions — argue that the symbol grounding problem is not a permanent wall but an engineering problem, and that they are solving it in pieces. A system that has sorted a million images of red has some tie between its word red and the look of red that a pure text model does not. Not a body's full lifetime of coping — no one is claiming that. The claim is smaller and harder to dismiss: grounding is not something a system simply has or lacks. It comes in degrees, and these systems have begun, partially, to acquire some.
And once that word — degrees — is in the room, the yes-or-no question begins to look strange. Consider what you have been demanding: a verdict, understands or does not, one or the other. But the thing you are measuring — grounding, contact with the world — turns out to come in amounts. So does almost everything else here. A whole line of current researchers says the all-or-nothing question is the mistake. There is a 2023 paper in PNAS by Melanie Mitchell and David Krakauer that does nothing but map this exact disagreement among the researchers, and it is explicitly unsettled — the experts are divided, and not for want of experiments. Their point, and the point of researchers like Ellie Pavlick, who go inside these models and probe what they actually represent, is that understanding was never a single switch. It comes in degrees and in kinds. A system can grasp the grammar of something and miss its point. It can track the inferences and miss the reference. It can have a great deal of one kind of understanding and none of another. Asking but does it really understand, yes or no is like asking whether a cluttered, half-sorted desk is really tidy — the question assumes a line where there is a gradient, and demanding the line is the error, not the answer. On this account both gut verdicts — obviously a parrot, obviously it understands — are malformed. They are answers to a badly built question.
There is a more radical version of this, worth hearing because it is the most thoroughgoing of them. Daniel Dennett spent a career on a position he called the intentional stance, set out in his 1987 book of that name: when you have something whose behaviour you can best predict by treating it as if it wants things and believes things and means things, then that is all there is to it. Treating it so, when it works, is the whole of its having a mind or meaning anything. There is no further hidden fact, no extra ingredient called real meaning sitting behind the successful interpretation, waiting to be checked. Donald Davidson reached a near neighbour of this from the side of interpretation, in his 1973 paper "Radical Interpretation": meaning just is what a careful interpreter, trying to make the best sense of you, is warranted in taking your words to mean. And if that is right, then your question — but does the machine really mean it, underneath? — is chasing a fact that was never there for anyone, yourself included. You engage with the thing all day by treating it as meaning things. It works. On this account there is no court of appeal beyond it works — no deeper fact about meaning you are failing to consult. The question yes, but really? has nothing left to answer to.
I want to mark the one genuinely balanced point in all this, because there is one idea that sits exactly between the two sides and will not come down on either, and it is the one we left unresolved last week. Recall where Wittgenstein left us: meaning lives in a practice, in being in the game with others, where you can be corrected and held to what you said. There is a developed theory of exactly that — Robert Brandom, in his 1994 book Making It Explicit, treats meaning as a matter of inferential role, your place in the game of giving and asking for reasons: to mean something is to be answerable for it, to owe reasons, to be entitled to conclusions, to be correctable by others who can say no, that does not follow. And this is the genuinely balanced point of the night, because it pulls equally in both directions. The machine demonstrably tracks the inferences — it moves from claim to claim in patterns that mostly hold, more reliably than you might expect. That pulls toward admitting it. But is it in the game, or has it only absorbed a record of the game? Is there anything it is committed to, anything it can be held to and be wrong about and answer for — or does it merely produce the next move with no stake in it? No one at present can say cleanly. Which is the point. The one idea that ought to decide the matter instead pulls both ways at once.
And beneath these dissolving arguments there is a quieter, more general thought supporting them — the most general of the pro-machine ideas, and one you have not yet been given a name for, so let me state it plainly. Perhaps what a word means, for any system, is not fixed by a body or a history or an intention at all. Perhaps it is fixed by what the word does inside the system — the job it performs, how it connects to everything else, how it drives what the system goes on to say and conclude and predict. Meaning as the work an inner state does within the whole machinery of thought. And if that is what meaning is, then whether the thing on your screen means anything is not a question to be settled from an armchair by checking it for a body or a soul. It is a question settled by going in and finding out what its inner states actually do — which is exactly what the researchers probing these models are trying to do. It turns the whole dispute from a verdict into an experiment.
Now here is the part you have to hold together, and it is the hardest of the night, because it would be easy to let one side cancel the other and be done. It does not cancel. Take both at once. The differences from the first half are real. The machine really does lack the history, the body, the reaching-for-you — those were not tricks, and these dissolving arguments refute none of them. What the dissolving arguments do is something else: they deny that any of those real differences adds up to a verdict. It lacks a body is true. It lacks a selection-history is true. Therefore it understands nothing, full stop — that is the inference they reject, because it assumes the very thing they deny: that understanding is the kind of all-or-nothing thing a single missing ingredient could switch off. So you do not get to collapse it either way. You cannot say the differences are real, so the verdict is in — because real differences do not make a verdict unless understanding is a switch, and that is exactly what is in dispute. And you cannot say the question is malformed, so the differences do not matter — because the differences are sitting right there, genuine and unrefuted. Two true things that do not reduce to one: there is a real difference, and is there a real difference, yes or no may be the wrong question. Accept both at once. Do not abandon either to make it easier.
Let me say plainly what changed tonight, because for once it is not simply another thing removed — you gained something.
You came in with a worry the last few weeks had been feeding: that perhaps there is nothing here at all — that the distinction you are so sure of is one you invented, and that you are no different from the machine in any way that counts. Tonight that worry was answered. There are differences, and they hold. The machine was never built to track anything; you come from a long line of creatures that were. It has no body that ever met the world; you have one that has touched water and lifted weights and pulled back from heat. There is no one within it reaching to be understood by you; you reach for other people all day. Those are real. You were not inventing the distinction. So take that — it is the gain, and it is clean: the worry was mistaken, and something is genuinely there.
And in the same breath you gained the harder thing, the one that keeps it honest. You learned that a real difference may not be what you should have been looking for. Because a second set of serious arguments granted you every one of those differences and then asked the question that removes the verdict: so what? Understanding may come in degrees and kinds rather than yes or no, in which case it lacks a body is true and therefore it understands nothing simply does not follow. Meaning may be no more than what a competent interpreter is warranted in taking a thing's words to mean — in which case but does it really? is a question with no fact beneath it. So here is the tool, and it is a strange one to hold: you must now accept two claims at once. There is a real difference between you and the machine. And is there a real difference may be the wrong question. An hour ago you would have said one of those must cancel the other. Tonight you can see that they do not, and you can stop looking for the one claim that would let you dismiss the other.
This is the fifth of nine, and the shape of the whole series is now visible. The first four weeks took the easy answers apart — the parrot, the company, the inner grasp — each of them coming back through your own words. Tonight went the other way: the differences that hold, and the arguments that say holding was never the point. You have now met the whole field. Every serious thing anyone has said about what meaning is — the accounts that keep the machine out, and the accounts that say the question is broken — all of it is in front of you.
Which is why next week stops being a survey. We have laid out everyone else's answer. There is only one question left worth asking, and it is not about the machine at all. Of all these — function, body, intention, use, company, degrees, the working stance — which one is yours? You have been relying on one of them this whole time, without saying so, to keep the machine on its side of the distinction. Next week I ask you to say which, out loud. For now, simply notice that you have been choosing — and that you have never once examined what you chose.
Before then, run tonight's work on a case I have not touched. You have heard me hold these two things together; now find out whether you can do it without me.
Here is the word: heavy. Specifically — a box at the bottom of the stairs, and the question of whether it is too heavy to carry up. Ask the machine. It will do the job well: it will reason about the mass, the grip, the angle of the stairs, the state of your back, whether to bend at the knees, when to ask for help. Genuinely useful. Now run tonight's first arguments on it. Was its word heavy ever built to track anything? No — it never carried a thing in its life, never had a body that strained, nothing in it was ever shaped by the actual cost of lifting. On that test there is a difference, and it holds: you know too heavy in your arms and your lower back, in a way it plainly does not, and that is not nothing. Now run the other arguments on the same word. Does it understand heavy — yes or no? Notice that the question resists a yes-or-no answer. It grasps more about heaviness than most people you know — the physics, the ergonomics, the risk — and it has never felt an ounce of it. So is that a yes or a no? It is neither. It is a thing that has a great deal of one kind of understanding of heavy and none of another kind, and the yes-or-no box has no slot for that. Hold the box at the bottom of the stairs in mind and try to deliver the verdict, and you will find a real difference and a question that will not take a one-word answer, both at once, with neither one letting you dismiss the other.
That is where tonight leaves you. You came in afraid the distinction was nothing. You leave knowing it is something — and no longer sure that something or nothing was ever the right way to ask.
Thank you for listening. I will see you next time.