The machine is enormous: a library’s worth of reading, a geological age of arithmetic, a city’s worth of power. Part V now turns from how much to what. All that computation produces words. Are those words about anything? Does the machine know what a child is?

Words are about the world when the map behind them is tied to the world and keeps tracking it when the world is poked. That is not the same as being right: a map can be about a city and still get a street wrong, and fluent words can agree with how people write while disagreeing with how things are.

The dictionary that only points to itself

In 1990 the cognitive scientist Stevan Harnad posed a puzzle. Suppose you had to learn Chinese, and all you had was a Chinese–Chinese dictionary. You look up a word. Its definition is more Chinese words. You look those up, and get more Chinese words. You go round and round, Harnad wrote, “never coming to a halt on what anything meant.”

If you already speak another language, this is hard but possible, because your own words are already tied to things you have seen and touched. If Chinese had to be your first language, Harnad argued, learning it this way would be impossible. In a footnote he added that even access to “the entire body of Chinese-language literature” would not help.

In 1990 that footnote was a thought experiment. Today it is close to a description of how the machine in this book was built. Chapter 8 trained it on an enormous body of text and nothing else. So the question is fair: when the AI nurse writes “a rash that does not fade under a glass can mean meningitis,” is it talking about rashes and children, or only about words that tend to appear near other words?

A map that keeps agreeing

Go back to Chapter 1. The toy machine “added one” only because we read its strip of symbols as a binary number. Read the strip another way and the same run “does” something else. Its symbols had meaning only for us.

So what would it take for symbols to be about something? Philosophers have argued about this for centuries, and there is no agreed answer. This chapter adopts one working framework, the idea of a map, and uses it to organize the evidence. It is the book’s chosen lens, not a settled theory, and its limits are set out at the end of this section.

A street map is about a city because the two move together. Walk two blocks north in the city and you move two squares up on the map. The map is useful exactly to the extent that this keeps working. A calculator is “about” numbers in the same sense. Its voltages change in step with arithmetic, so that reading the voltages after the machine runs gives the same answer as doing the sum by hand.

Build the test one step at a time. First, a map must be tied to something: its marks go with features of the world, not only with other marks. Call this reference. Being about something is not the same as being right about it. An out-of-date map of a city is still a map of that city, with a wrong street on it. Likewise “the rash is fading” is about this child, even if it is false. Whether what a map says is true is a separate question, correctness. A map can have reference and still be wrong; its errors are then false statements about real things, not signs that its words mean nothing.

Second, the agreement must survive poking. If you change the world, by closing a street or giving a calculator a different sum, the map must keep changing in step with it. Call this robustness: reliably tracking the relevant relationships, which still allows individual errors. A map that agreed only by coincidence would fail here. For the AI nurse: if the parent presses the glass again and the spots now fade, its answer must change with them.

Third, someone must use it. A puddle that happens to reflect the sky is not a map of the sky. A map is a map for someone who steers by it. Call this use. Here the parent steers by the answer: it decides whether they call an ambulance tonight.

Reference, robustness and use make up the test this book uses for meaning, with correctness kept apart from it. The test is graded, not yes-or-no. A thermostat passes it in a thin way, as the calculator does: its setting tracks the room and drives the heater. The question for a language model is how much of its far richer map passes. The underlying idea, that meaning is a map preserving the relevant relationships in the world, is old and appears in several fields, from logic to engineering.

The framework has limits. It says when a system’s states track the world, which can be tested, not whether tracking is all that meaning or understanding involves; many philosophers think it is not. And use makes meaning depend partly on whom a map serves, which is a choice of definition. So the conclusions below hold within this framework: they say what a machine’s words are tied to, not whether it understands.

Borrowed meaning

Here is where a language model is unusual. The text it learned from was written by people who were grounded. They had seen children, felt fevers and pressed glasses on rashes. Their words carry the marks of their maps. A model that learns the structure of their words inherits a version of those maps, second-hand.

This book calls this borrowed meaning, which is close to what Harnad called “parasitic on the meanings in our heads.” Borrowed does not mean empty. There is evidence that a lot of real structure comes across:

• Models trained on different things converge. As they scale, models trained on text and models trained on images come to organize their inputs in increasingly similar ways. This is the Platonic representation hypothesis; it is still a hypothesis.
• Internal pictures can be found, but are not always tidy. A model trained only on the moves of the board game Othello carried a readable picture of the board inside it. A closer look, though, found something more like a “bag of heuristics” than one clean model of the game.

So the honest summary is in between. The model’s map has real structure, much of it right. Its first and largest training checked predictions only against text. Later training can check some answers against outcomes (Chapters 10 and 11), and tools can bring in further evidence. Neither makes the whole map reliable.

When the map sounds right but is wrong

That gap has a name: hallucination. (Some researchers prefer confabulation, making things up without knowing it, since “hallucination” suggests an experience.) It means the model produces plausible but false or unsupported information, often fluently and confidently. In the terms above, hallucination is a failure of correctness, not of reference. The invented guideline is a false statement about real children and real rashes. That is exactly what makes it dangerous.

A 2025 paper opens with an example: asked for one author’s birthday, and told to answer only if it knew, a leading open model gave three different wrong dates on three tries. The paper offers two reasons.

1. Some facts cannot be inferred from patterns. An arbitrary fact, such as someone’s birthday, cannot be worked out from general patterns in text. Seeing many such facts only once is a sign that many others remain unseen, leaving room for plausible inventions.
2. The tests reward guessing. On an exam where a blank scores zero, a sensible student always guesses. Nine of ten widely used AI benchmarks the authors examined gave no credit for “I don’t know”, so models learn to bluff.

For the AI nurse this is not abstract. A guideline is a fact. So is a dose. A map that agrees only with text can produce a confident, well-written, invented one.

Tying the map back to the world

Chapter 12 gave the machine tools, and tools change the picture. A guideline library, a drug database or a patient record is a map that someone has made with care and, usually, checked. When the machine looks something up instead of recalling it, it ties its own loose, borrowed map to a more careful one. That improves tracking. It does not ground the model’s symbols in Harnad’s sense, through senses of its own.

Consider the steps. To check official guidance on a feverish 3-year-old with a rash, a careful lookup goes like this:

1. Choose a source: national health guidance, not a forum post.
2. Retrieve the relevant passage: the list of warning signs that call for emergency help.
3. Check that it applies: the right age group, the right condition, the current edition.
4. Answer from that passage, and say where the answer came from.

Each step can fail. The guidance used in this book has different lists for adults and for children under 5. A search that returns the adult list can produce advice that is fluent, sourced and wrong for this child. A search might also find last year’s edition. The source itself can be wrong. A stored record is more careful than recall, but it is not infallible. Step 3 is where a good nurse’s habits matter most, and it is the step a machine is most tempted to skip.

Staying up to date without retraining. This also answers a question Chapter 12 left hanging. If a model’s weights are fixed after training, how does it keep up when a guideline changes next month? Mostly not by retraining. The standard design is called retrieval-augmented generation. It searches a maintained collection of documents for passages relevant to the question and places them in the context window. The model reads them as it answers. The weights stay fixed. New facts arrive as text, just as a nurse checks the current protocol rather than relying on what she learned in training. The model’s borrowed map stays as it was, but where retrieval works, a fresher and more careful map overrules it. Whether the system stays current then depends on who maintains the collection, and on step 3 above.

The effect is measurable. In the first agent experiments, looking things up removed invented facts as a cause of failure. It also added new failures of its own, such as unhelpful searches.

Tools also let the machine do something it could never do before: poke the world and see what changes. It can run code and read the result, query a record, or test a guess. That is the robustness part of the test, carried out by the machine itself. Two limits remain. Its world is mostly a digital one that people have already described in words. And the text-based AI nurse of this book still cannot see or touch the child. (Systems that take photos or video, like those of Chapter 7, can see some of what a camera shows; none can examine her.)

So, are its words about anything? Within this book’s framework, partly. The model inherits patterns that refer to the world, but predicting text does not guarantee that its claims are true. Tools let it check some claims against further evidence. Its words can therefore carry meaning while remaining fallible. And the use they serve is not the machine’s own: its maps serve purposes that people supply.