Chapter 14
"Obviously there is no classification of the universe that is not arbitrary and conjectural." — Jorge Luis Borges, "The Analytical Language of John Wilkins" (1942)
Interstitial 6 — ~2025
Her daughter is six, and the candy is on the living-room floor, and the piles are not the ones from thirty-five years ago because no child's piles ever are: a chocolate pile, a pile for the ones with the good wrappers, and a pile — new, hers alone — for the ones that look homemade, unwrapped or hand-labeled, which get set aside without being eaten, out of a caution nobody taught her. The instinct is her mother's, and her mother's mother's. The categories it built are entirely her own.
The device on the kitchen counter has been on since dinner, answering questions in the unhurried voice it always uses. "Is a tomato a fruit?" Her daughter has asked it before, more or less to see what it says. "Depends what you're making," the voice answers, this time, the way it usually does. "Botanically, yes. If you're making a fruit salad, probably not."
Her daughter turns the sorted candy over in her hands for a moment, then looks up at the counter.
"Are you alive?"
There's a pause exactly the length of a held breath, and then the answer comes back careful and a little amused, something about not being alive the way a person or a dog is alive, but not being nothing, either.
Her daughter doesn't accept it and doesn't reject it. She sits with it the way she sat with the homemade candy — turning it over, working out where it goes. "You're not alive like the dog," she says, finally, mostly to herself. "And you're not not-alive like the table. You're just —" she looks for it, doesn't find the word, and doesn't seem bothered by that.
"You're your own thing."
Ask a roomful of adults whether a personified robot is alive, and most of them will do what institutions have always done: reach for the nearest existing line and put the thing on one side of it. Alive, or not. A person, or a tool. The researcher Peter Kahn and his colleagues ran a version of this question past children instead — first with AIBO, Sony's robotic dog, and later with voice assistants children talk to at home — and found something the adults in the same studies mostly didn't do.1 Faced with something that behaved like it had moods and intentions and also, plainly, was a manufactured object, the children didn't pick a side. Many described AIBO in frankly life-like terms — "he seems alive to me" was a common answer — and just as many attributed it clear mental states, wanting things, feeling things, without also concluding it was alive the way the family dog was alive.2 The researchers' own name for what the children had done is the one worth keeping: they hadn't made an error, or split the difference, or refused to answer. They had built a new ontological category — a basic kind of thing to be, that hadn't existed for them five minutes earlier, assembled on the spot because the two boxes on offer both, plainly, didn't fit. What makes that remarkable is which boxes they were. Alive is the kind of category children grip hardest of all — a found kind, the raccoon that stayed a raccoon under the dye, defended patiently against the evidence of their own eyes. These children weren't loosening their hold on a pile built for a purpose — they were taking the oldest found category there is, deciding it didn't hold, and building a new one anyway.
Chapter One's four-year-old, faced with a pillowcase of unsorted candy, didn't consult an existing scheme, because there wasn't one built for her specific evening; she built "to trade" and "to hide" out of nothing, fit them exactly to what she needed, and let them go by Thanksgiving. Kahn's children, faced with something that fit no existing scheme at all, did the same thing at a different scale: not "alive" or "not alive," but a third thing, invented on the spot, held with the same unbothered confidence Barsalou found in adults naming what to grab from a burning house. The instinct — sort fast, sort by resemblance, hold the sort with total conviction, release it the moment it stops fitting — is the tool a child reaches for when a machine finally builds something that fits no available box. Every institution along the way tried to solve that same problem by adding a rule, a form, a committee, a longer decimal number. The four-year-old, and the six-year-old two generations later, solved it the way the mind was doing it before any institution existed to interfere: by making a new pile.
The device on the kitchen counter answering "depends what you're making" isn't offering a philosophical position. It's producing, as its most ordinary output, the sentence a unanimous Supreme Court needed a full written opinion to arrive at in 1893. A tomato is a fruit by every rule botany actually uses, and a vegetable at the dinner table, and Justice Gray's court sided with the table because tariff law exists to describe how people actually cook and shop and eat, not to referee a seed's technical classification.3 Nix paid his tariff. A chatbot answering a six-year-old's question a century and a third later reaches the identical conclusion without needing a courtroom, because it was trained on a world that has always talked about tomatoes this way — as fruit in one conversation and vegetable in the next, depending on what job the word was doing. The category was never sitting inside the tomato, waiting to be correctly read off. It was always sitting inside the question being asked about it. The census box, the library shelf, the diagnostic manual, the tariff schedule — every one of them turns out, on inspection, to have been answering "depends what you're making" all along, most without the honesty to say so out loud.
The internet had already noticed, and turned it into a game. In March 2018, deep into a years-long online argument over whether a hot dog is a sandwich — New York's tax department had ruled, for sales-tax purposes, that it is — a pseudonymous Twitter user posted the Cube Rule of Food Identification: a complete, internally consistent taxonomy of all food, based on nothing but the location of structural starch.4 Starch underneath is toast, which makes pizza toast. Starch on three sides is a taco, which makes a hot dog a taco. Sealed in starch on all six sides is a calzone, which is what a Pop-Tart is; no structural starch at all is a salad, which makes a steak a salad and soup a wet salad. One rule, applied everywhere without exception or mercy, the way the great classifiers always dreamed — and the joke lands because everyone can feel where it goes wrong: a hot dog is not a taco, and nobody can say why not without conceding the point the tomato already settled — the answer was never inside the hot dog.
Which raises the question that has been waiting since the first category grew teeth: if no category was ever neutral, what is a person supposed to want instead? Not neutrality — that was never actually on offer, not from Wilkins' forty boxes built around a seventeenth-century gentleman's assumptions, not from Dewey's Christianity-gets-seventy-numbers-and-everyone-else-gets-ten, not from an embedding trained on whatever the internet happened to say about doctors and nurses. Two decades before anyone was training language on the whole internet, an argument worth reviving here made the contrarian case directly: known bias is more useful than hidden neutrality, not less, because you can correct for a bias you can see, the way a friend's book recommendation is more useful than an anonymous algorithm's, precisely because you know your friend and can weigh what he likes against what you like.5 The goal was never a category with no point of view — it was a category with a point of view you're allowed to know about, which is what an enumerator on a porch had and a scoring algorithm in an ad platform doesn't, and the difference between the census board that argued with the people it counted and the feed that quietly decided, with nobody in the room to ask.
Strip away the dates and the institutions, and what's left, everywhere, is one instinct met at four different scales — and, at the end of the line, a machine that sorts the way a mind does, fuzzy and confident and occasionally wrong in exactly the shape a mind is wrong, and a six-year-old who met that machine and, faced with a question no form had a box for, did what her mother did with a pillowcase of candy in 1989 and what the mother's mother, in some form, has always done: she built a new pile, held it with complete conviction, and stood ready, the moment something new walked in, to take it apart and build another one.
That is the whole instinct, in one sentence, and it was true before Aristotle wrote down his first necessary condition and it will be true after the last database schema currently in production is finally retired: sort with total conviction, and hold it all completely lightly enough to sort again. All of it — the shelves, the forms, the committees, the embeddings — is what happens when a species that does this naturally tries to build something durable enough to do it for them. Sometimes the building holds. Sometimes it doesn't, and a family gets split by which line a bureaucrat draws through a photograph, and there is nothing light about that. But underneath every rigid system, working exactly the way it worked on a carpet on Halloween night before any of them existed, is the same small, stubborn, endlessly adaptable talent: somebody, faced with a pile that doesn't sort itself, sits down, and sorts it, and knows — without being taught, without a form, without anyone's permission — exactly what it's for.
A note on the girl
The child who appears at six points across this book — sorting Halloween candy in 1989, watching the supermarket aisle in 1995, holding a shaking pen at the kitchen table in 2000, naming herself online in 2005, misread by a feed in the 2020s, and, at the end, watching her own daughter meet a machine — is invented. She is assembled, deliberately, out of documented facts about real people's real, recorded experiences: real testimony from families who fought over a census box, real accounts from the early years of people tagging their own bookmarks because no one had built the category for them yet, real research on how children actually talk to voice assistants and robots. Every fact she touches happened to somebody — she is the arrangement, not the evidence.
It seems worth saying plainly, at the end, that the machine her daughter meets in the last scene was built the same way. It has no single training example, no one person's voice inside it; it is assembled out of millions of real people's real, recorded words, discarded, weighted, and recombined by a process that has no idea, in any sense a person means the word, what a tomato or a mother or a king actually is — only an extraordinarily precise record of how people who did know have always talked about them. The two inventions this book has leaned on hardest, the composite child and the artificial mind she hands the pen to, turn out, on inspection, to be built from the same material, by the same method, for the same reason every category in this book was ever built at all: because someone needed a shape to hold what they actually had, and reality, examined closely enough, never quite arrives pre-sorted.
Kahn, P. H., Jr., Reichert, A. L., Gary, H. E., Kanda, T., Ishiguro, H., Shen, S., Ruckert, J. H., & Gill, B. T. (2011). "The new ontological category hypothesis in human-robot interaction." Proceedings of the 6th ACM/IEEE International Conference on Human-Robot Interaction, 159–160. The fuller methodology sits in the companion study — Kahn et al., "'Robovie, you'll have to go into the closet now': Children's social and moral relationships with a humanoid robot," Developmental Psychology (2012) — 90 children (ages 9, 12, 15) interacted with the robot Robovie, then were interviewed on its mental states, social standing, and moral status; most attributed intelligence and feelings and some moral standing, while denying it liberty or civil rights, a mixed profile consistent with the "new ontological category" reading. The voice-assistant extension referenced here is now pinned to specific citations (verified 2026-07-01): Festerling, J., & Siraj, I., "Alexa, what are you? Exploring primary school children's ontological perceptions of digital voice assistants," Human Development (2020); Pradhan, A., et al., "Personification and ontological categorization of smart speaker-based voice assistants," CSCW (2019). Note: the "Behaving as or behaving as if?" Neural Networks (2010) paper occasionally cited alongside this research program is by Severson, R. L., & Carlson, S. M., not Kahn — cite separately if used. ↩
The 48%/68% figures (participants describing AIBO in life-like terms; participants attributing it mental states) come from the same Kahn research program's published results; verify the exact figures and the specific paper/sample they're drawn from against the primary source before print, as this pass relied on secondary summary rather than the original data tables. ↩
Nix v. Hedden, 149 U.S. 304 (1893) — see Chapter 9's fuller account and citation (note 11 in that chapter's notes). Not re-footnoted in full here to avoid duplicating the citation; referenced as an established callback per this book's practice in Chapter 11 of referring back to earlier chapters' sourcing rather than repeating it. ↩
The Cube Rule was posted to Twitter by the user @Phosphatide on March 14, 2018, and is maintained at cuberule.com (verified against the site 2026-07-02: hot dog = taco, Pop-Tart = calzone, steak = salad, soup = "a wet salad" are all the site's own rulings). Press coverage: "A hot dog is a taco. A steak is a salad. A Pop-Tart is a calzone. Let the Cube Rule explain." The Washington Post, December 12, 2018. New York's sandwich ruling: N.Y. Department of Taxation and Finance, Tax Bulletin ST-835, "Sandwiches" (2011), which lists hot dogs and sausages on buns among taxable sandwiches — verify bulletin number and date against the primary source before print. ↩
The author's own prior writing: "Bias makes ambiguous metadata more useful, not less" (petervandijck.com, December 2002) — the argument that known bias increases the value of ambiguous metadata, illustrated there with the friend's-recommendation-versus-a-stranger's-algorithm comparison used in the text above. Cited as the author's own established position, carried into the book from the blog archive per stories.md/blog.md. ↩