What the Machine Can’t Carry

I was sweating through my shirt before I even reached the podium. My hands were shaking so badly I could see it. I'd pictured this moment a hundred times in the weeks leading up to it, and none of those pictures looked like what was actually happening to my body as I stood there. I was a graduate teaching assistant, and I'd never taught at that level before, except I had taught my daughters to walk, and nearly everything else that comes with raising small children, which is its own kind of patient, repetitive instruction, just never at this scale or in front of strangers. I was there to talk about the Black Death. Still, only because I wanted it to be about something else: the kids in that room who had lived through their own pandemic, who had their own memories of empty streets and closed schools and the particular loneliness of watching the world go muted. The professor I worked with had them share those memories out loud, so that by the time I reached 1348, the room was already primed to feel 2020 sitting right next to it.

A chatbot could have given those students the mortality rate, the trade routes, the theological crisis the plague provoked more accurately than I did, shaking the whole time. What it could not have done was need them to understand it. I needed them to, because I had my own memory of a muted world sitting right alongside theirs, and some part of me was asking them to witness that, not just learn from it. The explaining was never the point. The collapsing was the moment two separate griefs, six hundred years apart, recognized each other in the same room. That requires someone present, with something at stake in whether the recognition actually happens. AI can produce the lecture. It cannot want anyone to believe it.

That need is the part I can't stop thinking about, because it's also the part a machine cannot inherit, no matter how much of our history it's given to hold. A machine can hold a family recipe, a country's founding, and a stranger's diary entry in the same way: as data. A few lines of text, tokenized, weighted, sitting beside billions of other lines of text with no more or less claim on its attention than anything else. The recipe doesn't know it's sacred to anyone. The founding document doesn't know it cost anyone their life. The diary entry doesn't know it was never meant to be read by a stranger. It takes a person to decide that one of those things mattered more than another, to decide, the way I decided to make a plague lecture about a pandemic those students had actually lived through, and that decision, small and constant and usually invisible, is most of what we actually mean when we say "memory."

Inheritance has never really been a transcript. It's always been a choice, made by someone who had something to lose or protect in the choosing: a parent deciding which version of a story to pass down, a nation choosing its textbooks, a historian deciding what belonged in the record and what didn't. AI doesn't choose like that, because it has nothing to lose. What it has instead is a staggering quantity of other people's choices, their stories, their records, their softened and hardened and unrecognizable versions of events compressed into something that can produce a plausible next sentence. In a very real sense, it is inheriting everything and nothing at the same time. Everything, because the material is all there, somewhere in the weights. Nothing, because none of it cost it anything to receive.

I think about this in terms of who used to decide what got passed down to me, and who decides now. It used to be people and a nation choosing its textbooks, a church deciding which parts of its own history were worth repeating aloud, an archive deciding whose letters were worth keeping and whose were thrown out with the rest of the estate. Those choices were often wrong, sometimes badly, and sometimes on purpose. Entire histories have been curated into silence this way not erased exactly, just never selected, never digitized, never translated, never considered important enough to keep. I spend a lot of my own research thinking about that kind of exclusion: how the record of a thing can survive in full and still manage to leave almost everyone out of it. AI doesn't invent that problem. It inherits it, at scale, and then quietly launders it into something that looks neutral. A gap in a national archive reads as an absence. A gap in training data reads as an answer, delivered in the same confident sentence as everything else, with no seam showing where the silence used to be. The new curator doesn't have skin in anything. It has whatever was most common, most published, most said, and that becomes the version that survives not because anyone decided it should, but because no one decided anything at all. The worry is that AI will not get history wrong. People have always gotten history wrong, usually on purpose, usually for a reason you could at least argue with. The worry is that it has no reason to get it right, and no reason to get it wrong either. No one is home to hold responsible for the version it hands you, and nothing to appeal to when the version is missing something you needed.

Which brings me to the question I keep circling, the one I don't think I can fully answer: if a machine can produce something that reads as knowledge, or even as narrative, without ever having inherited anything itself no fear, no family, no stake in getting it right then what does making something still mean? I used to think authorship was about the words landing correctly, in the right order, saying the true thing. Increasingly, that was always the smaller half of it. The larger half was choosing those words while risking being wrong, being judged, being misunderstood, being seen. I feel that risk every time I publish an essay here under my own name, not a pseudonym, nothing to hide behind; my husband reads it, my daughters will be able to read it one day, people I actually know in real life read it and then see me out in the real world afterward. That exposure is not incidental to the writing. It's what makes the writing mean anything at all. A parent softening a story for their children takes the same kind of risk, deciding what version of themselves they want remembered alongside the story, knowing it might be wrong, knowing it will eventually be judged by the very people it was meant to protect. AI takes no such risk. It has nothing to protect, no shoes to put on to walk the same steps as you afterward, no children who will someday read what it said and hold it against the person who said it. Nothing, in the deepest sense, is on the line, which means nothing, in the deepest sense, is being authored. Only produced.

That could be the real inheritance question of this moment, bigger than whether AI gets the facts right. Not what will it remember, but what will it cost it to remember wrong, and the honest answer is nothing, which means it was never really remembering at all. I keep returning to that distinction because it's the only place I can locate what's actually missing, the thing I can't quite name otherwise. It was doing something that looks like memory from a distance, the way a photograph looks like a moment until you realize the moment itself the temperature of the room, who flinched, who reached for whose hand is the one thing the photograph can never give back to you. A person who misremembers their own history loses something real: a relationship, a reputation, a piece of self-understanding they have to live inside thereafter. A machine that misremembers loses nothing, because there was never a self there to live inside the consequence. That's what unsettles me more than any specific factual error it might make. It isn't the errors. It's that the errors cost it absolutely nothing, which means that somewhere underneath all the fluent language, no one is actually keeping score.

I don't know what to do with that, except keep choosing to tell the story myself, inadequately and differently depending on who's in the room, the way people always have. Not because I'll get it more right than a machine could; I almost certainly won't, not on the facts, because I'll have something at stake in getting it wrong. Because someone will read it and know I decided what to leave in. Because the risk is the authorship, and I'd rather keep both than hand either one away.