Chapter 13

Updated 2026-07-02

"We say 'the portrait resembles the person' rather than 'the person resembles the portrait.'" — Amos Tversky, "Features of Similarity" (1977)

In July 1945, with the war in the Pacific still unfinished, an engineer named Vannevar Bush published an essay in The Atlantic that named, more precisely than anyone had before him, the constraint every shelf, census form, and filing cabinet had been quietly fighting all along. Bush had spent the previous several years directing the government office that coordinated some six thousand American scientists toward the war effort — not a librarian, not a philosopher, the man who had just finished running the country's wartime research.¹ And what he chose to write about, the moment the fighting allowed him to look up, was the ordinary, unglamorous problem of finding things again. "Our ineptitude in getting at the record is largely caused by the artificiality of systems of indexing," he wrote. A fact, a book, a record, filed by subject, "can be in only one place, unless duplicates are used… one has to emerge from the system and re-enter on a new path."²

One place. It's Dewey's rule stated as a complaint rather than a design — one book, one shelf, one number, and no way to hold two true addresses at once. It's the census box, forcing one parent onto the line where two belong. It's Ranganathan's whole objection to the tradition he'd been trained in, arrived at independently, on the other side of the world, by a man who'd never read Ranganathan and never needed to, because the constraint was the same constraint everywhere a shelf, a form, or a filing cabinet was involved. Bush's fix, though, wasn't a new set of facets but a different foundation entirely. "The human mind," he wrote, "operates by association" — not by indexing, not by drawers and subheadings, but by leaping from one thing to a thing it resembles or borders or reminds you of. He proposed a desk-sized machine, the memex, that would let its owner build "trails" of association between any two documents, so that "any item can be joined into numerous trails" instead of exactly one.³ The memex was never built — the mechanism Bush imagined was mechanical and photographic, overtaken within a decade by the digital computer — but the essay is credited, by the engineers who spent the next several decades building hypertext and the early Web, as the place the idea started.⁴ Bush had drawn the map in 1945. It took most of a century for a machine to actually walk it.

That machine arrived, in a modest and almost accidental form, in 2013, when a team of Google researchers published a method for representing words as points in space.⁵ Feed a computer enough ordinary text — news articles, by the billions of words — and have it learn, purely from which words tend to appear near which other words, a location for every word in a few hundred dimensions nobody could draw or visualize directly. Words that get used in similar contexts end up near each other. And then the researchers noticed something almost too clean to be believed: take the location of "king," subtract the location of "man," add the location of "woman," and the point you land on, in that invisible space, sits closest to "queen." No one told the machine what a king or a queen is. No one gave it a rule, a definition, a family tree of royal succession. It had only ever been shown examples — millions of them, in context, the way a child is only ever shown examples of dogs before the word "dog" starts to mean something — and out of nothing but the accumulated pattern of those examples, it had located something that behaved, mathematically, exactly like the concept of gender crossed with the concept of rank.⁶

This isn't a database. A database with a table called "royalty" and a column called "gender" would get to "queen" instantly and for a boring reason: someone built it in on purpose. Word2vec was never told any of that. It arrived at a structure that looks, from the outside, like Rosch's prototypes and Lakoff's radial categories — fuzzy, similarity-based, learned from examples rather than declared by rule — because that shape falls naturally out of predicting text, the same way it falls naturally out of a four-year-old's brain sorting Halloween candy. Human categories, Rosch's work showed, are neighborhoods, not boxes: a center, graded membership, resemblance without a single defining rule. An embedding is a neighborhood built out of nothing but co-occurring words, and the neighborhoods it builds keep landing in the same places human intuition would put them. One honest caveat, and the field itself insists on it: the clean version of the king/queen example only works if you throw out "king" itself as a candidate answer before asking what's nearest; leave it in, and the point closest to king−man+woman is very often just king again.⁷ The machine isn't discovering eternal truths about royalty and gender — it's finding the strongest pattern in the text it was shown, and the strongest pattern, examined too literally, sometimes just says: more of the same. Held with total conviction, and just as ready to be re-sorted — the same instinct that sorted a pillowcase of Halloween candy, now running on a billion documents a machine was never asked to understand, only to predict.

Nowhere is that same machinery visible at a more absurd scale than in the back rooms of a streaming service trying to get you to watch something. In January 2014, the journalist Alexis Madrigal noticed that the genre labels Netflix used internally — things like "Understated Independent Dramas Based on Books" or "Emotional Fight-the-System Documentaries" — were reachable through sequentially numbered web addresses, and wrote a script that spent twenty straight hours walking through every number.⁸ It surfaced 76,897 of them. Traced back to its source, the system ran on Ranganathan's idea at industrial scale — facets, recombined: paid contractors watched entire films and rated them, by hand, on dozens of independent dimensions — how gory, how romantic, what decade, what setting, what ending — and a machine recombined those facets automatically into label after label after label, more of them than any single person would ever bother to name on purpose. "Romantic" alone showed up as a qualifier in 5,272 separate categories.⁹ Nobody sat down one morning and designed "Deep Sea Horror Movies from the 1970s." The facets existed; the machine did the combining; the category assembled itself out of parts a human being had, in fact, sorted one at a time.

Spotify ran the same experiment in reverse, and one man's career shows what happens when the human half of that arrangement quietly stops being necessary. For years, a data alchemist named Glenn McDonald sat between Spotify's genre-clustering algorithms — which sorted the platform's music into thousands of statistically distinct clusters no person had proposed — and the public, giving each cluster a name evocative enough that a listener could tell, from the label alone, roughly what it sounded like: "escape room," "stomp and holler," dozens of regional and decade-specific splinters of reggae and folk and hip-hop that had never had names before because nobody had ever needed to distinguish them until a machine found the gap between them first.¹⁰ By the time Spotify laid off fifteen hundred employees in December 2023, McDonald among them, his site Every Noise at Once had named 6,291 genres across roughly a million artists — fifty-six kinds of reggae, two hundred and two kinds of folk, two hundred and thirty kinds of hip-hop.¹¹ The layoff cut off his access to the underlying data. The site simply stopped updating, mid-genre, the way a card catalog stops updating the day nobody's left to file the new acquisitions. McDonald had spent a career standing at the exact point where a machine's raw clustering met a human being's need for a word to call it by. The machine didn't fire him. The company that owned the machine decided, one afternoon, that the point where a human stood between the clustering and the public was an expense rather than a service — and a man who had spent years handing names to categories a machine found first learned, the way a filing clerk might, that the machine no longer needed anyone standing at that particular desk.

None of this should read as the machine arriving at some cleaner, more neutral version of categorizing than the human systems it learned from. It inherited the mess along with the method. Word embeddings trained on ordinary news text learned, unprompted, that "doctor" sits nearer "man" and "nurse" sits nearer "woman," the same lopsided association Chapter Three's idealized cognitive models would have predicted, now baked directly into the geometry a search engine or a résumé-screening tool relies on.¹² Researchers have built methods to detect and partially correct this, which is itself a strange kind of progress: teaching a machine to notice its own prototype effects, the way people can be taught to notice theirs. And at the largest scale of all, in content moderation, the same machinery is asked to sort billions of individual pieces of speech into categories like "hate speech" or "harmless joke" continuously, worldwide, with no version of the task that satisfies everyone reviewing the results — an observation sharp enough that people who study the field simply call it an impossibility theorem and move on to arguing about which failures are tolerable.¹³ It is the ethnic-affinity problem again, at its hardest: a category with real consequences, assigned at a scale no human institution has ever attempted, by a system that is, in the end, still just finding neighborhoods in space.

There is one place, though, where the machine still visibly diverges from the mind it's imitating, and it was proven decades before the machine existed to test it against. In 1977, the psychologist Amos Tversky showed that human similarity judgments aren't symmetric.¹⁴ We say the portrait resembles the person, never the person resembles the portrait. We rate North Korea as more similar to China than China is to North Korea, because we measure the unusual thing against the familiar one, the variant against the prototype, and direction matters to that measurement in a way it has no reason to if similarity were simply a distance, the same both ways round. An embedding's similarity is a distance — cosine similarity, symmetric by construction, the space between two points measured identically no matter which one you start from. So an embedding can approximate almost everything about human category structure: the neighborhoods, the graded membership, the king-to-queen arithmetic. It cannot represent the one fact about human similarity Tversky proved forty years before anyone had a machine to check it against — that when we compare two things, we are never quite standing in the middle. We're always standing on one side, looking at the other, deciding which one gets to be the example everything else is measured from.

That's the seam. The machines took the pen, learned to sort the way minds have always sorted — fuzzy, prototype-based, built from examples instead of rules — and inherited nearly everything about the instinct except the one small, human habit of picking a side to stand on. And by the time they had, something was already sitting on kitchen counters, answering questions in an unhurried voice: a thing that fits no neighborhood at all, no cluster, no trail — a thing nobody, human or machine, had ever seen an example of before.

---