“We think that knowledge and understanding belong to art rather than to experience, and this because the former know the cause, but the latter do not.” (Aristotle, Metaphysics)
An AI system can pass a medical licensing exam. It can diagnose rare diseases from a list of symptoms, recommend treatments, and explain its reasoning in fluent prose. By any functional measure, it knows medicine.
But Aristotle would have disagreed. He drew a hard line between experience and knowledge. The experienced doctor knows that a drug cures a disease. The knowledgeable doctor knows why. Aristotle’s example in the Metaphysics is specific: knowing that “when Callias was ill of this disease, this did him good, and similarly in the case of Socrates” is experience. Knowing why the treatment works for people of a certain constitution is knowledge. Experience tracks particulars. Knowledge grasps causes.
His test was simple and brutal: the person who knows can teach. Not recite, not generate fluent explanations, but transmit the principle from which the particulars follow. The person who merely has experience cannot teach, because they do not possess the principle. They only possess the cases.
By this standard, AI may be the most experienced entity in history, and one of the least knowledgeable.
This distinction feels old. It is old. But it turns out to be the sharpest diagnostic tool we have for understanding what AI does when it produces a correct answer. Because AI produces correct answers constantly, fluently, confidently. The question nobody pauses long enough to ask is: does getting the right answer mean you know something?
The question “what counts as knowledge?” has been fought over for twenty-four centuries, and every major answer turns out to be relevant to what we are building now.
Plato, in the Theaetetus, explored and rejected several candidates. Knowledge is not just perception. It is not just true belief. His dialogue ends in aporia, in honest inability to resolve the question, but along the way he establishes something crucial: whatever knowledge is, it requires logos, an account, a reason. True opinion without logos is unstable. It “runs away,” as Plato puts it, because there is nothing anchoring it to the truth. A system that produces true outputs without possessing any account of why they are true has, at best, true opinion. Useful, but not knowledge.
Aristotle took the next step. Knowledge, for him, is specifically knowledge of causes. The senses, he writes, “do not tell us the ‘why’ of anything, for example, why fire is hot; they only say that it is hot.” Sensation tells you what. Experience tells you what, reliably, across many cases. Knowledge tells you why. And the hierarchy is precise: experience arises from many memories of the same thing; from experience comes art and knowledge, because the person with knowledge grasps the universal cause rather than merely accumulating particulars.
This framework held, with variations, for nearly two thousand years. The medieval scholastics added that knowledge requires grasping the essence, the formal cause. To know fire is not just to know it is hot. It is to understand what makes fire fire.
Then Hume detonated the structure.
Hume’s argument is simple and devastating. We never actually perceive causation. We perceive constant conjunction: A followed by B, repeatedly. We see the billiard ball strike, and we see the second ball move. We never see the causing. Causation, Hume argued, is a habit of expectation. We observe A followed by B enough times, and we start expecting B whenever we see A. That is all. As he put it:
“All inferences from experience, therefore, are effects of custom, not of reasoning.” (David Hume, Enquiry Concerning Human Understanding)
This matters enormously for AI, because Hume just described, with eerie precision, what a neural network does. A system trained on massive data learns that certain inputs reliably precede certain outputs. It builds associations based on constant conjunction. It does not perceive causes. It perceives patterns. Hume’s account of human induction turns out to be, 250 years early, an almost exact description of machine learning’s operating principle. The parallel is not a metaphor. It is a structural identity. Both Hume’s mind and a neural network learn by observing regularities and forming expectations. Neither has access to the causal mechanism that produces the regularity. Both, as a result, are vulnerable to the same failure: when the pattern breaks, neither can explain why, because neither ever knew why the pattern held.
Kant saw the problem and tried to save knowledge by adding structure. Raw experience is not knowledge. The mind must actively organize experience through categories: causation, substance, unity. These categories are not learned from data. They are the preconditions of having any experience at all. Knowledge, for Kant, requires structured understanding, not mere pattern extraction. The question this raises for AI is precise: does a trained model bring structure to data, or does it merely extract structure from data? There is a real difference, and it matters.
In 1963, Edmund Gettier showed in three pages that even justified true belief is not sufficient for knowledge. You can have a belief that is true and justified, and still not have knowledge, because the justification and the truth are connected by accident. This turns out to describe a common AI failure mode: producing outputs that are true and supported by training data, but where the connection between the support and the truth is coincidental rather than causal. Hallucinations that happen to be true are paradigmatic Gettier cases.
In the previous essay, we established that AI’s fundamental material is representation. Every input, operation, and output is a sign pointing at something else. The system never touches the thing itself. Now we can ask what a system made entirely of representations can know.
Here is what I want to argue: AI has experience without knowledge. It has prediction without understanding. It occupies a specific cognitive level that Aristotle identified twenty-four centuries ago and called empeiria, and it occupies that level at a scale no human being has ever approached. That is both its extraordinary power and its fundamental limit.
The most common objection to this view comes from reliabilism, the position in contemporary epistemology associated with Alvin Goldman: a belief counts as knowledge if it is produced by a reliable process. By this standard, a well-trained model that consistently produces true outputs might qualify. But reliabilism has a problem that bites especially hard for AI: the generality problem. Is the relevant process “GPT answering questions,” “GPT answering medical questions,” or “GPT answering this specific question about this specific patient”? Reliability rates diverge dramatically across levels of description, and there is no principled way to fix the grain. More fundamentally, Aristotle already anticipated this objection. He fully grants that the experienced doctor reliably cures. That reliability is what makes experience valuable. But reliability is compatible with complete ignorance of causes. The experienced doctor and the knowledgeable doctor may have identical success rates. The difference is that one can teach, and the other cannot.
If knowledge requires knowing causes, AI fails straightforwardly. When an AI system tells you that water boils at 100 degrees Celsius, it does not know why water boils. It does not know what boiling is. It knows that the tokens “water,” “boils,” and “100 degrees” co-occur with high probability. It has Aristotle’s experience without Aristotle’s knowledge.
Judea Pearl, the computer scientist and philosopher of causation, has arrived at essentially the same diagnosis from the engineering side. Pearl’s “ladder of causation” distinguishes three levels: association (seeing patterns), intervention (doing things to test causes), and counterfactual reasoning (imagining what would have happened otherwise). Current AI operates almost exclusively on rung one. It sees. It associates. It does not intervene, and it does not reason counterfactually. Pearl’s formulation is blunt: “All the impressive achievements of deep learning amount to just curve fitting.” The convergence with Aristotle is not a coincidence. Both are drawing the same line between tracking what happens and understanding why. Rung one of Pearl’s ladder is empeiria. Rungs two and three are episteme. The framework is twenty-four centuries old and still the most precise diagnostic we have.
Yann LeCun has reached a similar conclusion from within deep learning itself. LLMs, he argues, lack configurable predictive world models. They predict tokens, not states of the world. They have no capacity for the structured, causal, counterfactual reasoning that would constitute genuine understanding. That one of the architects of modern deep learning sees this need confirms the diagnosis from the inside.
The strongest opposing view belongs to Ilya Sutskever: “Predicting the next token well means that you understand the underlying reality that led to the creation of that token.” Recent mechanistic interpretability work has given this claim empirical weight. LLMs develop internal representations that track real-world structure in surprisingly sophisticated ways. This is a genuine challenge. But internal representations that track structure are necessary and not sufficient for knowledge. Aristotle’s test is not whether you have internal states that correlate with reality. It is whether you can give the logos, articulate the cause, and teach. A circuit that tracks a pattern is more evidence of sophisticated experience, not evidence of knowledge.
Murray Shanahan has put it most clearly: LLMs “cannot participate fully in the human language game of truth.” Emily Bender and Alexander Koller, in their “octopus test,” showed that a system trained only on the form of language, without access to its meaning, cannot genuinely understand what it produces. Contemporary epistemologists working on the concept of understanding, particularly Stephen Grimm and Duncan Pritchard, have given this ancient distinction its modern analytic form. Understanding, they argue, is a distinct cognitive achievement from merely knowing facts. It requires grasping causal and modal structure, the ability to explain, draw consequences, and apply principles to novel cases. Grimm traces this explicitly back to Aristotle’s “knowing-why.” This is not an antiquarian framework. It is live philosophy. And it describes precisely what AI lacks.
The diagnosis, as before, does not stop at AI.
Elizabeth Fricker, in a recent paper on AI testimony, has identified something that cuts deeper than the question of whether AI “understands.” AI systems, she argues, cannot transmit knowledge because they cannot take epistemic responsibility for their assertions. A human expert who tells you something stakes their credibility, their career, their reputation on the claim. An LLM stakes nothing. It produces a sequence of probable tokens. This is not testimony. It is simulation of testimony. And if we treat it as real testimony, we are making a category error with practical consequences, because testimony without responsibility has no epistemic weight, however fluent it sounds.
This ought to trouble us, because it does not only apply to AI. Most of what any individual “knows” was not discovered firsthand. It was absorbed from books, lectures, conversations, and screens. We hold beliefs about evolution, quantum mechanics, and the structure of the economy that we cannot independently verify and whose causal mechanisms we cannot articulate beyond a rough sketch. We trust that experts know the causes. We operate on their testimony. In a meaningful sense, much of what we call knowledge is really trust in other people’s knowledge, mediated by representation.
Hume’s description of induction as nothing more than the habit of associating constantly conjoined events was meant as a skeptical challenge to human knowledge. It turns out to be a more accurate description of machine learning than of human cognition. But it is not wholly wrong about us either.
A student who passes an exam by memorizing well has demonstrated experience, not knowledge. A professional who speaks fluently about a field without grasping its foundations has sophisticated experience, not knowledge. A citizen who forms confident opinions by absorbing the statistical regularities of their media environment has beliefs, possibly true, possibly justified, but rarely knowledge in any sense Aristotle or Plato would recognize.
AI did not create this condition. The gap between what we call knowledge and what knowledge actually requires has been widening since Hume. But AI has made the gap visible by building a system that operates entirely on the experience side of Aristotle’s distinction, and showing how far experience alone can take you. The answer is: very far. Far enough to pass for knowledge in almost every functional context. Far enough to make us wonder whether most of what we do is closer to experience than we would like to admit.
The question this leaves us with is not whether AI knows anything. It is whether we know as much as we think we do, and what it would take to genuinely know rather than merely to predict correctly.
We have identified what AI is made of (representation) and what it can know (experience without understanding). The harder questions keep coming: whether AI has purpose, whether it thinks, whether it can create. Each of these questions has its own history, and each one cuts differently.
We will continue to take them one at a time.
Minerva