What Is Knowledge?

You believe a lot of things. You believe the sun will rise tomorrow, that antibiotics treat bacterial infections, that your friend is honest. But which of those beliefs count as knowledge? That question — deceptively simple, philosophically explosive — is the opening problem of epistemology, the branch of philosophy that studies the nature, sources, and limits of human knowing. It has been asked for twenty-five centuries, and AI has made it newly urgent.

The Classical Account: Justified True Belief

The dominant answer in Western philosophy for most of its history is Justified True Belief (JTB). On this account, you know that P if and only if three conditions hold simultaneously: First, P is true. You cannot know something false. If you believe the Earth is 6,000 years old, that belief may be sincere, but since it is false, it is not knowledge. Second, you believe P. Knowledge requires mental commitment. A fact you have never encountered cannot be something you know. Third, your belief is justified. You have good reasons for believing P — reasons that are connected to what makes P true. A lucky guess that turns out to be correct does not qualify. All three conditions are necessary; any one alone is insufficient. Truth without belief is just a fact you have not encountered. Belief without truth is error. Belief without justification is a hunch. Only the intersection of all three — a true belief that you hold for good reasons — reaches the bar we call knowledge.

The JTB account was the standard until 1963, when philosopher Edmund Gettier published a three-page paper that shook epistemology. Gettier constructed cases where all three conditions are satisfied yet knowledge seems clearly absent. A classic variant: You glance at a clock on the wall and it reads 3:15. You form the belief 'It is 3:15.' It happens to be exactly 3:15 — your belief is true. And you were justified in trusting the clock, which has always been reliable. But unknown to you, the clock stopped exactly 12 hours ago. Your belief is justified and true — but it seems wrong to say you knew it was 3:15; you got lucky. Gettier cases reveal that justification plus truth is not sufficient for knowledge. You need justification that is connected to the truth in the right way — not coincidentally, not accidentally. Philosophers have proposed many refinements (reliabilism, virtue epistemology, safety conditions), but no universally accepted fix exists. The problem of what it really means to know remains open.

Types of Knowledge and Why They Matter

Knowledge is not a single thing. Philosophers distinguish several varieties: Propositional knowledge (knowing that): knowledge of facts expressed as propositions — 'I know that Paris is the capital of France.' This is the form most analyzed in epistemology and most relevant to the JTB debate. Procedural knowledge (knowing how): knowledge of how to do something — how to ride a bicycle, how to write a persuasive essay, how to debug code. You can know how to swim without being able to articulate any of the propositional facts about buoyancy that underlie swimming. Acquaintance knowledge (knowing of): direct familiarity with something through experience — 'I know what grief feels like,' 'I know Paris, I have walked every arrondissement.' This is direct, non-propositional contact with a subject. These distinctions matter in the age of AI because large language models appear to have extensive propositional knowledge and can even simulate procedural knowledge — but their relationship to acquaintance knowledge is contested. An AI system has never felt grief, never been cold, never experienced the texture of a rough wall. Whether this absence limits what it can truly know is not a trivial question.

Knowledge and AI: A Strange Mirror

When you ask a language model 'What is the capital of France?' and it answers 'Paris,' has it expressed knowledge? Let us apply the JTB framework seriously. Is the output true? Yes, Paris is the capital of France. Does the model believe it? Here things get complicated. The model does not have beliefs in the sense of mental states with subjective commitment. It has patterns in weights that generate certain outputs in response to certain inputs. Is the output justified? The model produces 'Paris' because of statistical patterns in training data. But justification in the epistemological sense requires reasons that are connected to what makes the claim true — not just correlational patterns. Most philosophers would say that on the JTB account, AI systems do not have knowledge — they have something that mimics propositional knowledge at the output level while lacking the mental states the account requires. But this raises a hard question: if an AI system reliably produces true outputs with a process connected to the truth (training on accurate sources), does the label 'not knowledge' matter? The practical stakes of that question grow every day.