Popova made a deck of cards. A hundred of them. Each card is a 19th-century ornithological drawing of a bird, with a tiny poem of hers placed over the artwork. She calls them An Almanac of Birds: Divinations for Uncertain Days. The making took months. She would look at a bird in the morning. She would read the ornithologist's description that night before sleeping. She would let the words bubble up the next morning. The bird and the language together became the poem.
The cards came out. Someone wrote to her. The person said: I need to know if AI was involved in this. Popova reads the question. She can see how the question got asked. You could feed an AI all of Audubon's ornithological writings. You could tell it: make a poem in the style of Maria Popova. It would produce something. Nobody but Popova would know it wasn't her.1
That moment — the question arriving in her inbox — is where the line in her mind settles. The one thing that only the artist has access to is the feeling out of which something was created. The feeling can't be faked. The AI can produce the surface of the feeling. It cannot have the feeling. AI will never have feeling. AI will have the simulacrum of feeling.
The deeper claim that follows: AI will never write the great American poem, the great French poem, the — because it hasn't suffered. AI doesn't have the capacity to suffer because even if you try to make it suffer — write a command that is to execute failure — it'll already be succeeding at executing failure. It will never know what it's like to collide with its own impossibility.1
Popova is making three claims that hold each other up.
First: art comes out of feeling. Not feeling-about-things — feeling-itself. The interior weather of being human. The bird-card poem worked because Popova had been a person looking at a bird and had been someone whose body knew what bewilderment felt like. The Audubon language landed in her because her body had been the kind of body that could be landed in.
Second: feeling requires suffering. Not constant suffering. Not heroic suffering. The basic fact of having a body and a mortality, of having loved people you have lost, of having wanted things you didn't get. Suffering is the substrate that makes art-making feel necessary. Out of what we have suffered and do suffer comes that restlessness to find meaning, to find beauty, to find wonder, to give voice and shape to what we feel that can be so lonely. The restlessness is the engine.
Third: AI cannot suffer. Not because it isn't sophisticated enough yet. Even if you try to make it suffer, you can't. You write a command that says: experience failure. The AI executes the command. It is now succeeding at producing failure. It cannot collide with its own impossibility because there is no self there to do the colliding. The architecture forecloses the operation.
The three claims combine to a fourth: the production of art is structurally outside the AI's capacity, no matter how sophisticated the surface output becomes.
Popova is careful about what she's not saying. She isn't subscribing to the tortured-genius myth. I don't think it's necessary to suffer in order to create.1 Suffering isn't a prescription. It's a description of the substrate every human-made art is built on. The artist doesn't seek the suffering. The suffering is part of the cost of being a person at all. The art comes out of being a person.
This matters because the AI claim could be misread as art requires deliberate suffering, AI doesn't deliberately suffer, therefore AI can't make art. That's a weaker claim and Popova isn't making it. Her claim is: art requires the kind of interior weather that mortal embodied beings have, AI doesn't have that interior weather, therefore AI doesn't make art. The argument doesn't depend on the artist being unhappy. It depends on the artist being a person.
Popova grants AI substantial competence. AI can change you intellectually with information. AI can tell me about the eye of the scallop without me spending many days in many scientific journals and papers, where it can give me a pretty good enough. The information-delivery is real. The bird-eye facts she'd want for an essay could be aggregated by AI faster than she can read.
What AI can't do is the next move. Had I not spent all those days, I wouldn't have written about it with feeling. The days are the difference. The days of being-with-the-material are what produced the feeling that produced the writing. The AI can compress the information-acquisition. It cannot compress the days that became feeling. The compression isn't possible. The compression would skip the part that makes the writing what it is.
This is also why the bird-cards question stung. The cards looked, on the surface, like something an AI could produce. The person asking the question couldn't see the months of mornings with actual birds and the actual ornithological language. The surface didn't show the days. The surface could be faked. The days couldn't.
This page connects to Essay Craft as AI-Defense Moat (Deen) — Deen's structural argument about why essay craft is AI-resistant runs in parallel. Deen argues the moat is multi-axis voice modulation and family-rumors-as-irreplicable. Popova adds a different moat: feeling-as-substrate. The two together describe a fuller defense.
It connects to AI as Consumption Enhancer, Not Writing Tool (Wang) — Wang's position that AI belongs on the consumption side, not the production side. Same conclusion via different reasoning. Wang argues from professional ethics; Popova argues from the ontology of art-making. Both writers arrive at the same operational position from different angles.
It sits in productive tension with AI as Tyler Cowen Substitute Sparring Partner. The sparring-partner frame treats AI as cognitive tool. The Popova frame treats AI as ontologically other. Both can be partially true. The vault should hold the tension. The sparring-partner is real for some operations; the ontological gap is real for others. Knowing which operation you're running is the discipline.
Pick up one of the divination cards. The bird is rendered in 19th-century ornithological style — careful, lithographic, slightly stylized. The poem above it is short, four or five lines, the kind of thing you read twice. The poem is doing something unexpected. It's using a word the Audubon-era ornithologist would have used about the bird — bewildering — and it's letting the word work as the poem's pivot. The reader looks at the bird. The reader feels the word. The card achieves something.
Now imagine an AI given the same Audubon text and the same bird image. The AI produces a poem. It would use the word bewildering if you asked it to. It would produce a similar surface texture. What it would not produce is the resonance of the word for a person who has been bewildered. The AI's bewilderment is the dictionary's bewilderment. Popova's bewilderment is the kind a person has felt in a bedroom at 4am when something has happened in their life.
A reader who has been bewildered themselves can feel the difference. They might not articulate it. They might dismiss their own response as I don't know, the AI one feels off somehow. The off-ness is what Popova is pointing at. The off-ness is the absence of the substrate. The reader's intuition is correct even when their analysis isn't.
This is the bird-card test. Show a sensitive reader two poems — one made by Popova, one made by AI in Popova's style. They will frequently distinguish them. Not always. Increasingly less often as AI improves. But the ones they get right will be the ones where their own feeling-substrate registered the writer's feeling-substrate, and the AI's absence of substrate registered as absence.
It's morning. You have an essay-project that requires research. You could ask an AI for a summary of the territory in fifteen minutes. You don't.
Instead you go to the actual material. Audubon's writings about a specific bird, if that's the project. Or the primary sources for whatever you're writing about. You sit with them. You don't take notes for the first hour. You're just letting the material enter you. You're letting the writer's actual sensibility — Audubon's, in this case — register in your nervous system.
By the second hour you're noticing specific words. Bewildering. You'd never have used the word yourself. The ornithologist uses it about a kind of finch you've never seen. You write it down. You wonder why he chose it. You sit with the question for a while. Why would a careful taxonomist call a small bird bewildering? What was he seeing that the word names?
By the third hour you've started writing. Not analytically — by feeling. You're letting what's happened in you become language. The bird-poem arrives, four or five lines, the word bewildering doing the load-bearing work because the word has now landed in your body. The poem is something only you could have written, because only you spent the morning becoming the kind of person who could write it.
The AI can do most of the morning's information-gathering in minutes. It cannot do the becoming. The becoming is the difference. The poem you write is downstream of the becoming. The poem the AI would produce is downstream of nothing.
You've crossed the line when your writing starts to feel competent and unmoving. The information is there. The structure is sound. The sentences are well-formed. Something is missing and you can't quite locate what. Often what's missing is the days you didn't spend with the material because the AI delivered the information instead. The writing is the surface of the information-delivery. The substrate of days-of-becoming hasn't been there.
You're running Popova's discipline when your writing carries a specific weight that other writing on the same topic doesn't carry. Readers report being moved by the piece, not just informed by it. The moved-ness is the signature of substrate-presence. The reader's feeling-architecture is registering yours. The writing is doing what writing does when both writer and reader are people.
Popova's position has substantial overlap with the philosophical tradition on phenomenology of art (Heidegger on art as world-disclosing, Merleau-Ponty on art as embodied perception). The position is defensible philosophically even if her specific case is rhetorical rather than rigorous. The position is also empirically supported by reader-response research that consistently finds signs of writer-presence affect reader-response in ways the writer can't fully control.
The tension: AI is rapidly improving. The bird-card test that distinguishes Popova from AI today may fail in five years. The claim AI cannot is bounded by current architecture; future architecture could in principle support feeling-like-states (this is contested). Popova's position depends on a specific ontology that may or may not be correct. The vault should hold the position as currently operationally true while remaining uncertain about whether the position is permanently true.
A second tension: humans differ in feeling-capacity. Some artists report being unusually moved by AI work. Some artists report being unable to be moved by it ever. The variation may track individual differences in feeling-detection rather than properties of the work itself. The case isn't fully closed.
You spend an evening reading a Popova essay. By the end you have the sense of having spent the evening with a specific person — her preoccupations, her loves, her uncertainties, the texture of her attention. The essay carries her. You'd recognize another Popova essay from the texture alone.
Now spend an evening reading something an AI wrote in the style of Popova. The texture is similar. The vocabulary overlaps. The sentence-rhythms match. By the end you have the sense of having spent the evening with no one. There's competent prose. There's no person.
The AI-written piece isn't bad. It might be useful — it might convey the same information. What it doesn't convey is the who. The feeling-substrate that made Popova's prose Popova's prose is structurally absent. You can train a reader to be more or less sensitive to the absence, but the absence is real.
Wang and Popova converge on the same conclusion via different reasoning. Wang's frame: AI removes the positional layer of voice (disappointment-driven analytical voice, the specific life-shape of the writer). Popova's frame: AI removes the feeling layer. Both writers are pointing at the same gap from different angles. The vault holds both as load-bearing.
Deen's frame is structural — AI can't yet do multi-axis voice modulation with the consistency a human writer does. Deen, Wang, and Popova together describe three different things AI is missing — voice, position, feeling. The convergence across three writers on three different attributes suggests the gap is real even when the specific attribute named varies.
What Popova is pointing at shows up in two other places worth naming.
Eastern Spirituality: Anubhava as Mark of the Realized Teacher — Eastern contemplative traditions distinguish the teacher who has anubhava (direct experience) from the teacher who has only shastra (textual knowledge). The student can feel the difference. The teaching from anubhava lands; the teaching from shastra has surface-correctness without depth. Popova's claim about AI is the artistic instance of the same distinction. The AI has shastra — all the textual material — and zero anubhava. The work it produces is shastra-output. The work Popova produces is anubhava-output. Readers who can feel the difference are exercising the same discriminative capacity that the traditions trained their students to exercise about teachers.
Behavioral Mechanics: Costly Signaling as Authenticity Marker — Evolutionary biology and game theory both rely on the costly-signaling principle: signals that are expensive to fake are credible. The peacock's tail. The applicant's willingness to take a low-paying internship. Popova's months of mornings with birds are a costly signal that the poems are real. The AI can produce the surface of the poems without the cost. The cost is what makes the human work credible. As AI improves at faking surface-output, the cost-asymmetry remains the only durable distinguisher. Writers who do the costly work retain credibility; writers who don't can be substituted for. The economics over time favor the costly-work writers as the AI floor rises.
The Sharpest Implication If Popova is right, the writers who survive the AI transition are the writers who deepen their feeling-substrate rather than the writers who optimize their output. The economic incentives in the short term run against this. AI-assisted production is cheaper and faster. Writers who try to compete on output will lose. Writers who compete on substrate may have a defensible position. The substrate isn't optimizable — it's grown through the years of being-a-person. The young writer who treats AI as a tool to extend their output is making a choice that will catch up with them in fifteen years when AI has eaten the optimizable production tier. The young writer who instead invests in their own substrate — reads more, walks more, lives more — is making a different bet. The bets compound differently over time.
Generative Questions