Two things were said at the very end of last week almost in passing, and left unanswered on purpose. Tonight we return to both, because on inspection neither is small.
The first. When you finally stated your rule out loud — the machine does not understand its words because they are not connected to anything real — I said we would come back and ask where that idea came from. It did not come from nowhere. Early in this series you reached for a principle — that the meaning of a word is its connection to a thing in the world — and treated it as the plainest fact available. Tonight I want to show that it is not a fact you discovered. It was handed to you. It has a history. Someone built it, and in becoming the obvious view it displaced an alternative.
The second. The machine is fluent, and that fluency is what set this whole question going. But tonight I want to ask the question that sounds too simple to be worth asking: fluent in what? "Language" — the thing we keep asking whether the machine understands — was never a single thing sitting out in the world. It is always someone's.
This is Philosophy for Us — philosophy for everyone, no degree required.
This is the eighth of nine. Last week was the most demanding episode: you turned your own rule on your own words and watched it do two things at once — apply straight through some of your vocabulary and fail to apply to the rest — and you could not say which reading was correct. That episode is behind us, and it was the most difficult part of the series so far. Tonight is less demanding. But it does something the demanding episode could not: it examines the rule you were using and asks where you got it, and then it asks whose words the machine is actually speaking. Two questions you would have said had obvious answers. Neither does.
Start with the rule itself, because you have to see it clearly before you can see that it has a history. The one most listeners reached for: a word means something because it is connected to a thing. "Tree" means tree because it points at trees. To understand a word is to have made that connection — the word on one side, the thing in the world on the other, and the meaning is the relation between them. That does not feel like a theory. It feels like a plain description of how words work. Hold on to that feeling, because it is exactly what I want to call into question.
Here is the first surprise. The picture is old, and we can watch someone write it down. About fifteen hundred years ago Augustine, in the Confessions, described how he learned to talk as a small child, and what he wrote is almost word for word your rule. The adults, he said, would make a sound and at the same time turn toward a thing; he saw this and understood that the sound was the name of the thing; and little by little he collected these until he had enough to say what he wanted. Word, thing, the name stands for it. That is the whole account — your rule, fifteen centuries early.
You could read that and simply nod: yes, obviously, that is how it works; what else would it be? That nod is the point. It was obvious to Augustine, who wrote it down as a plain memory, and it was obvious to you at the start of this series, when you reached for it and treated it as certain. Fifteen hundred years apart, the same picture, and neither of you stopped to ask whether it was the only one available. That is not a fact about how meaning works. It is an inheritance.
Now watch what happened to the picture next, because Locke's refinement of it is the part that matters most tonight. A few hundred years ago John Locke, in the Essay Concerning Human Understanding, went further: when the word points at the thing, where exactly does the pointing happen? His answer, which became most people's answer, was that it happens in the mind. The word does not attach directly to the tree out in the field. It attaches to your idea of a tree — the concept you carry inside — and that inner idea is what the word actually stands for. He moved meaning inward: out of the world, into a private mental content the word is tied to.
Notice how much of your own reaction runs on that exact version. When you say the machine does not really understand — that it merely arranges the words with nothing underneath — what is the "underneath" you are reaching for? It is the inner idea: the thing in the mind the word is tied to, which you have and the machine lacks. That is not an observation you made about the machine last month. It is a three-hundred-year-old answer to a question Locke asked, passed down through everyone in between, and placed in your hands so early that you took it for your own perception.
So do not try to repair it yet. Just notice where you are standing. The rule you reached for is not the foundation it appears to be. It is the latest in a long line of received views, each handed on as simply the way things are. You did not discover it. You inherited it. And no one ever told you where it came from, or what it cost, or whether the people who built it would still stand behind it. They would not. And that is where it breaks down.
So the picture is old, it is inherited, and it was sharpened. Here is why that matters tonight — and it is not a history lesson. It is the thing that has quietly been causing your difficulty for six weeks.
The picture reached its high point about a hundred years ago, when the most careful people who ever worked on language tried to make it precise. Even there, two different projects were running that we should keep apart, because they are often blurred together and they are not the same. Gottlob Frege, in his 1892 paper "On Sense and Reference," argued that a word carries two things, not one: the thing it points at, and the way it presents that thing. "The morning star" and "the evening star" turn out to point at the same planet, Venus — but they plainly do not mean the same thing, because they present it differently. So meaning is not just the object at the end of the word; it is the object together with the way it is presented. That is one project: describing the word-to-world relation more honestly. The other project ran the opposite way. The logical atomists — Bertrand Russell, and the early Wittgenstein of the Tractatus — imagined a perfect language, in which every simple thing in the world had exactly one clean name, an exact map of word onto world. Keep the two apart: one says every name comes with a way of presenting its object; the other wants names so exact they would need none. But notice what they share. Both are still inside the old picture — word, world, and the relation between them. They are refining the inheritance. Neither is leaving it.
And then the picture broke down — not from outside, from sceptics, but from the people who held it most carefully and tested it until it failed in their hands. And here is the part to hear tonight: it did not break into two. It broke into a whole family of successors that contradict each other — and you have been relying on one of them, without knowing, as though it were the whole family.
Let me set the successors out, because you walked straight through the middle of this disagreement last week without knowing it was a disagreement.
One set of successors kept the core of the old picture — a word connects to a thing — but denied that the connection is in the mind, where Locke had put it. It is out in the world: a chain of causal contact running back to the thing itself, a chain you mostly never traced yourself. That is the water example from last week, from Putnam. Your word "water" means the substance here because of a connection out in the world, not because of any picture in your mind. On this view, meaning stays a connection to a thing; it simply stops being anything you privately own.
A second set of successors did something more drastic: they threw out the core. Naming a thing, they said — this is the later Wittgenstein — is just one of the many things we do with language, one activity among a great many, and the old picture's whole mistake was to take that single activity, pointing and naming, and treat it as the secret of all meaning. Look at what you actually do with words all day, they said. You are hardly ever naming. You are asking, joking, promising, counting, warning, comforting, cursing. If that is right, then "connection to a thing" is not the foundation of meaning at all. It is one activity among a hundred, and the old picture mistook it for the whole.
And a third set — the ones you met earlier — kept a real word-to-world connection but understood it as something the first two never mentioned: as a biological function, something a system was built over millions of years to track; or as a body that genuinely made contact with the world. And those successors, remember, do not admit the machine. They hand the difference back to you and let you keep it.
Now stand back and look at what you are holding. Four successors of one broken picture, and they deliver four different verdicts on the same machine. The causal-chain view partly clears the machine and partly clears you — on that view you are both borrowing your connections. The language-games view says there was never an "underneath" for the machine to lack, or for you to have. And the biological-function view and the embodiment view say the machine really is missing something you plainly have. The same family; verdicts that flatly contradict one another.
So here is what the history was for — the point of going back fifteen hundred years. Your rule, "its words are not connected to anything real," is not one clear idea with one meaning. It is a fragment of an old, broken picture, and the moment you try to make it precise enough to use, you are forced to choose which successor you meant. And you already saw last week what choosing does: some successors run the rule straight through your own words, and some let you keep it. The difficulty the machine has been causing was never really about the machine. It is a disagreement inside the idea you brought in — a disagreement that began a hundred years ago, among people who could not settle it, and you walked into the middle of it early on not knowing there were sides, holding one of them, and calling the one you happened to hold the obvious truth.
Set the history aside for a moment, because the other thing I left unanswered last week is a different kind of difficulty — simpler than the genealogy, and perhaps, by the end, worse.
The machine is fluent. That fluency is what drives all of this; it is what set two opposed reactions running in your head in the first place. So step back and ask the question that sounds almost too simple to bother with: fluent in what?
We have spent eight weeks asking whether the machine understands language — as though there were a single thing called language, sitting out in the world, and the only open question were whether the machine reaches it. But there is no single thing. What the machine was trained on was an enormous quantity of someone's words: written down, by particular people, in particular places, in particular centuries — far more of some kinds of people than others. More of those who wrote than of those who did not. More of the printed than the spoken. More of the recent than the ancient; more of some languages and some settings and some registers than others; and almost nothing from the people who were never written down at all. Its fluency is not fluency in the world. It is fluency in a record — and the record has a shape, and the shape is someone's.
We have seen this before in this series. Earlier we discussed knowledge, and who gets to decide what counts as knowing. There is no view from nowhere. Every account of the world is an account from somewhere, by someone, leaving something out. And here is a machine that is fluent in someone's portion of the world, presented to you as fluent in the world itself.
And this bears directly on the question we have been examining for two months. Earlier in this series, one of the most serious answers on the table was this: meaning is use. A word means what it does in the lives of the people who use it, within the whole form of life it belongs to. Take that seriously and turn it on the machine: whose form of life is it fluent in? It has none of its own. It never lived anywhere. It was never in the practice — never corrected across a life, never held to its words over years, never had to answer for one of them to someone who would remember. It is fluent in the records of our forms of life: an average of a great many of them, and not one of them its own.
So even the answer most favourable to the machine — the one that said competent use is all meaning ever was, with no inner extra beneath it — even that one, tested here, does not give the machine a clean pass. It hands you a stranger question instead. Can you be fluent in a form of life you were never in? Can you be a master of the use of words within a practice you never stood in, and only read the records of?
I am not going to answer that tonight. I am putting it in your hands, as with the others. But notice what it did to the question you came in with. "Does it understand language?" felt flat, neutral — one clear question with one obvious shape. And it was concealing a prior question the whole time: whose language, whose use. You cannot even ask the large question honestly until you have asked the small one you passed straight over.
Let me bring the two together, because they run parallel, and the parallel is what you take away.
Both this week and last, you found the same structure at work. Last week: the rule you had been using to measure the machine turned out to be something you inherited and never inspected — one successor of a broken old picture, reached for without examination, with several other successors you had never met standing right behind it. And tonight: the "language" you had been asking the machine to understand turned out to be someone's — weighted, partial, a record with a shape — while you had been treating it as the neutral whole of the world. The same structure both times. You took something as simply given — the obvious idea of what meaning is; language itself — and found it had a source, a history, a someone, that you had never thought to check.
Let me be careful here, because there is a cheap version of tonight I do not want you to leave with. The cheap version says: so it is all just inherited pictures and someone's records, none of it is real, and there is no difference between you and the machine after all. No. That is not what tonight showed. Some of those successors — the biological-function one, the embodiment one — still hand you a real difference and let you keep it. Finding out that your idea of meaning has a history does not flatten everything to one level. It does something more uncomfortable: it tells you that you cannot say which difference is the real one without choosing a successor, and you have never yet chosen on the merits. The ground did not disappear. You simply found out that you had never examined your claim to it.
So here is what you can take with you, and it is sharper than it sounds. From now on, when you feel yourself about to deliver the verdict — it obviously does not understand language — you have two questions you did not have a month ago. First: where did I get this rule I am about to apply, and did I choose it on the merits, or just reach for the nearest one? Second: whose language am I asking about — the world's, or one particular part of it, presented as the world's? That is not a technique for winning arguments. It is the difference between holding an idea and being held by one you never examined.
Try it before next week, on something small and entirely your own. Take a word you use constantly and treat as flat — "natural," say; or "normal"; or "fair." Run both questions on it, with no one prompting you. First: where did that meaning come from — what did it displace to become the obvious one — and would it still look obvious from somewhere else, to someone else? Second: whose use of that word have you been quietly treating as simply the meaning of it? You will feel the same thing give way that gave way tonight — the settled sense that the meaning was simply there, plain, yours, from nowhere. Do it once, all the way down, on one word.
This was the eighth of nine; one remains. Next week is the final episode, and every voice comes into it at once: the parrot, the use theorist, the people who trace the connection back out into the world, the ones who say a real difference still stands, and the ones who say the whole yes-or-no question is the wrong shape to begin with. All of them, together, with your own rule held up in the middle of them. And the question you came in with eight weeks ago — is it a parrot, or does it actually understand — is handed back to you, in front of all of them, sharper than you have held it before. That is next time.
Here is where tonight leaves you, and I am not going to smooth it over. You came into this sure of two flat, obvious things: that you know what your own words mean, and that the machine is fluent in language. Tonight both grew a history beneath them that you had never seen. The idea of meaning you were standing on turned out to be one side of a hundred-year-old disagreement you had joined without knowing. And "language" turned out to be someone's, not the world's. You cannot un-see either one now. The next time the machine explains something back to you — more clearly than you could have, with the one example you needed — and you reach for well, it does not really understand the language, the two questions will start up in your mind. Which idea of meaning am I using. And whose language did I just call the language. You will not quiet them again — and that difficulty is the whole of what the hour was for.
Thank you for listening. I will see you next time.