We use the word constantly.
“The AI knows your preferences.” “It knows the codebase.” “It knows how to write in your style.” We’ve imported the word so casually into our descriptions of these systems that we’ve stopped noticing what the word was doing before we borrowed it.
So let’s ask the question directly.
What does it mean to know something?
Not to have access to information about it. Not to be able to generate fluent text concerning it. Not to produce outputs that a knowledgeable person would recognize as accurate. Those are all real capabilities and they matter. But they’re not knowing. They’re something adjacent to knowing that we’ve been calling knowing because we don’t have a better word ready and the distinction didn’t used to matter practically.
It matters now.
Knowing in the Flores sense, in the Heideggerian, Maturana-grounded, background-constituted sense the entire series has been building toward, is not a state of having information. It’s a capacity. A structural readiness. The ability to navigate novel situations, cope with breakdown, make judgments at the edges where the rules run out, and do all of this from a background that was built through time and cannot be shortcut.
Knowing is what you have after ten thousand breakdowns have shaped your structure. After thrownness and embodiment and accumulated failure have deposited something in you that isn’t representable but is unmistakably real.
The question for AI is not whether it can produce knowledgeable outputs. It demonstrably can. The question is whether it can have that. Whether something like that is even possible for a system constituted the way these systems are constituted.
We don’t know. But we can be precise about what the question actually is.
Let’s map what knowing requires against what current AI systems have and don’t have.
Knowing requires structural history. A system that has been changed by its encounters with the world. Not updated, changed. The difference matters. An update adds information to an existing structure. A change in the relevant sense alters the structure itself. The background isn’t a database that gets new entries. It’s a living organization that gets reshaped by breakdown. Current AI systems are updated. Whether they can be genuinely changed in this sense is an open question that nobody is asking clearly because the word “learning” is doing ambiguous work across both meanings.
Knowing requires stakes. The background develops under conditions where something matters. Where failure has consequences that the failing system has to cope with. Where the breakdown isn’t just a signal to be processed but a perturbation that must be survived and integrated. A chess engine that loses a game loses nothing. A chess player who loses a game loses something, confidence, ranking, the psychological coherence of their self-image as a player, and the recovery from that loss deposits something in their background that the engine cannot accumulate. Current AI systems have no stakes. Nothing they generate has ever cost them anything in the sense that matters.
Knowing requires embodiment. Not metaphorically. Literally. A body that exists in a world that pushes back. That has a perspective, a here, a now, a set of concerns arising from biological existence. Maturana was insistent on this. Cognition isn’t implemented in nervous systems. It is nervous systems maintaining structural coherence in environments that resist. You cannot have knowing without the kind of structural organization that biological embodiment produces. Current AI systems are not embodied. They process. The question of whether processing can produce something functionally equivalent to embodied knowing is genuinely open, but it’s not being asked with the precision it requires.
Knowing requires commitment. The ability to make promises that bind you across time. To have an identity that is constituted by the network of commitments you inhabit. To be someone who can be held accountable, who has something at stake in the fulfillment of what was promised. Knowing without commitment is information without ground. You can know facts about surgery without being someone who operates on people whose lives depend on your judgment. The surgeon knows differently than the student who has read the same textbook. The difference is commitment. Current AI systems cannot make commitments in this sense. They can produce commitment-shaped language. The substance is absent.
Knowing requires the possibility of genuine breakdown. The moment when the current interpretive framework proves inadequate and must be reconstructed from a deeper background. Error is correctable within a framework. Breakdown requires a new framework. Current AI systems produce errors. Whether they can experience breakdown in the generative sense, whether they can be genuinely thrown by a situation that exceeds their categories and forced to reconstruct their interpretive ground, is perhaps the deepest open question in the field. And it is almost never asked this way.
So what would it take?
It would take a system that accumulates structural history through genuine stakes. That has something like embodiment, not biological embodiment necessarily, but the functional equivalent of existing in a world that resists and costs. That can make commitments and be changed by breaking them. That can experience breakdown as a generative disruption rather than processing it as an error signal.
Whether that’s possible is unknown. Whether it’s what anyone is currently building toward is easier to answer.
It isn’t.
This matters in two directions simultaneously.
The first direction is toward AI development. The field is making extraordinary progress on capabilities, on what these systems can produce, predict, generate, optimize. The Flores framework suggests that capability without knowing is a specific kind of danger. Not a malevolent danger. Structurally dangerous. The confident wrongness at the edges. The invisible colonization of the background. The formalization trap closes around systems and users simultaneously.
If knowing is the actual target, if what we want from AI is not just capable outputs but something like genuine judgment at the edges, then the current development trajectory may be running fast in a direction that gets further from that target with every increment of capability. More capability with the same structural absence of stakes, embodiment, commitment, and genuine breakdown is not progress toward knowing. It’s progress toward more fluent not-knowing.
The second direction is toward you. Right now. Deciding how to inhabit this moment.
The knowing gap is real. It’s opening. The background that constitutes genuine expertise is not being built at the rate it used to be built, in the domains where AI generation is replacing the productive friction that built it. That’s a civilizational fact and individuals can’t reverse it.
But individuals can decide what they’re building in themselves. Can choose deliberately to stay in contact with breakdown. To debug without assistance sometimes. To build things from scratch not because AI can’t build them but because the building is what builds you. To treat knowing as the actual asset rather than the output as the actual asset.
The output depreciates. Knowing compounds.
Every framework in this series, ready-to-hand, breakdown, background, speech acts, commitment, formalization, has been pointing at the same thing from a different angle. The thing that makes humans irreplaceable isn’t what we can produce. It’s what we’ve become through the process of producing it.
That’s what knowing is. That’s what’s at stake.
@ Lee Greene