What Is Epistemology?

A Working Definition

Epistemology is the philosophical study of knowledge, belief, and justification. It asks three central questions:

1. What is knowledge? What conditions must be satisfied for a person to count as knowing that some proposition is true?
2. What justifies belief? What does it take for a belief to be rationally held, and what is the relationship between justification, evidence, and truth?
3. What can be known? Are there propositions about which knowledge is possible, propositions about which it is not, and propositions where the question itself is contested?

The first question is conceptual: what does the word knowledge pick out? The second is normative: what makes belief well-grounded? The third is de facto: which beliefs survive the standards the second question imposes?

Epistemology connects directly to What Is Logic? and Validity vs Soundness: logic gives us the form of inference, while epistemology asks when inference plus evidence amounts to knowledge. It also connects to What Is Philosophy? since asking what counts as a good answer to any claim is itself the epistemic question. The discipline is one of philosophy's four traditional core areas, alongside metaphysics, ethics, and logic.

The Central Question

The central question of modern epistemology, in its sharpest form, is this:

Given that I have a belief $p$ that happens to be true, what additional condition must hold for me to know $p$ rather than merely believe-truly $p$?

Knowledge, since at least Plato's Theaetetus, has been treated as more than mere true belief. A lucky guess that happens to land on a true proposition does not count as knowledge. Whatever knowledge is, it must distinguish the lucky-guesser from the genuine knower. The history of epistemology is the history of attempts to specify the additional condition.

The classical analysis: knowledge equals justified true belief (JTB). To know $p$, three conditions must hold:

1. $p$ is true.
2. The agent believes $p$.
3. The agent is justified in believing $p$.

This was the working definition from Plato through the early twentieth century. Edmund Gettier's three-page paper in 1963 showed that JTB is insufficient: there are cases satisfying all three conditions where we hesitate to attribute knowledge.¹ The problem he raised, and the responses that followed, are covered in detail at The Gettier Problem. The brief version: any analysis of knowledge that uses justification in roughly the classical sense is vulnerable to constructed cases where the agent reaches a true belief through justified-but-coincidentally-correct reasoning.

Why Epistemology Matters

Three concrete cases.

Case 1: a clinical diagnosis. A patient has unusual symptoms. The attending physician orders a test, the test returns positive, and the physician issues a diagnosis. A medical AI system, trained on hospital records, returns a different diagnosis with a confidence score. The case forces an epistemic question: which assessment is justified? Both have evidence. Both have track records. They disagree. Epistemology gives us the conceptual framework to ask which evidence carries more weight, and why.

Case 2: a courtroom. Conviction requires belief beyond a reasonable doubt. What standard of justification does that name? Reasonable doubt is not a probability threshold, not a frequentist calibration target, and not a Bayesian credence cutoff in any direct sense. It is a normative epistemic standard whose specification is contested in legal theory and whose application varies by jurisdiction. The standard is doing real epistemic work, but no satisfying philosophical analysis of it has emerged.

Case 3: machine-generated assertions. A language model produces a confident assertion about a piece of code, a historical event, or a scientific result. The assertion is true. Should we say the model knows the result, believes it, or merely outputs a statistically likely token sequence? The question is not idle. The answer determines what kind of accountability we attach to the assertion, what kind of correction the system needs when it fails, and whether the system can be deferred to as a source. This is the synthesis question taken up at Knowledge, Justification, and LLMs.

In each case, the practical question is not solved by more evidence. It is solved by clarity about what evidence is and what relation it bears to belief and to truth. That is the epistemic task.

The Major Positions

Modern epistemology splits along several independent axes.

Foundationalism vs Coherentism (the structure of justification)

How is justification structured? Two positions:

Foundationalism holds that some beliefs are basic: justified non-inferentially, by direct experience, by self-evidence, or by some other immediate warrant. All other justified beliefs derive their justification from these basic beliefs, either directly or through a chain. The structure of justification is a tree rooted in foundations. Descartes is the canonical historical proponent (the cogito as foundational); among modern analytic philosophers, BonJour 1985 and Audi 2010 are well-developed defenses.²

Coherentism denies that any beliefs are foundational. Instead, justification is a property of the whole web of belief: a belief is justified to the extent that it coheres with the rest of the agent's beliefs. The structure is not a tree but a holistic system. Quine and Davidson are influential proponents in the analytic tradition.

The two positions face dual challenges. Foundationalism faces Agrippa's trilemma (covered briefly below): why are the foundations themselves not in need of justification? Coherentism faces the isolation objection: a perfectly coherent web of beliefs may still be wholly disconnected from reality. Most contemporary epistemologists hold mixed positions, allowing some beliefs more weight than others without insisting on a strict tree structure.

Internalism vs Externalism (where justification lives)

Does what justifies a belief depend only on what is going on inside the agent's head, or also on facts about the external world?

Internalism holds that justification is an internal matter: an agent and an exact mental duplicate of the agent must have the same justified beliefs. What justifies belief is some combination of the agent's evidence, reasoning, and access to her own mental states. Descartes is canonically internalist; Chisholm 1989 is a modern reference.

Externalism holds that justification depends on the agent's relation to the world, including facts the agent is not aware of. The most influential externalist position is reliabilism: a belief is justified if it is produced by a reliable belief-forming process, regardless of whether the agent knows the process is reliable. Goldman 1979 is the founding paper.³

The internalism / externalism split is one of the deepest in modern epistemology. It bears directly on AI: an LLM has no introspective access to whether its training was reliable, but its outputs may be more or less reliable as a function of facts about the model. If reliabilism is right, the model's claims may be justified in the externalist sense. If internalism is right, justification requires a kind of self-knowledge the model arguably lacks.

Rationalism vs Empiricism (the source of knowledge)

Where does knowledge come from? This split is older than the others.

Rationalism (Plato, Descartes, Spinoza, Leibniz) holds that some substantive knowledge can be acquired by reason alone, independent of sensory experience. Mathematical knowledge is the canonical example: $7 + 5 = 12$ is known not by observation but by reasoning.

Empiricism (Locke, Berkeley, Hume) holds that all substantive knowledge comes from sensory experience. Innate ideas are denied; the mind is shaped by what the world impresses on it.

Kant's synthesis in the Critique of Pure Reason (1781) argues that knowledge requires both: sensory content (the empiricist insight) and structural form imposed by the mind (the rationalist insight). Modern philosophy of science continues this discussion under different vocabulary: what is the role of theory-laden observation, of prior structure, of empirical confirmation? See the existing essay Empiricism, Induction, and the Limits of LLM Generalization for the rationalist-empiricist split applied to current AI.

Skepticism

Throughout the history of epistemology runs a parallel track: the skeptic, who argues that the standard for knowledge is so high (or that the obstacles are so severe) that little, if anything, can be known.

The two canonical skeptical arguments:

Cartesian skepticism (Descartes 1641, First Meditation). I might be dreaming, or being deceived by an evil demon. In either case, my beliefs about the external world would be exactly as they are. So I cannot tell from the inside whether I have any knowledge of the external world.

Pyrrhonian / Agrippan skepticism (ancient, restated by modern epistemologists). Any putative justification for a belief either rests on further beliefs (pushing the question back), terminates in dogma (an unjustified foundation), or is circular. None of these is satisfactory. Therefore no belief is fully justified.

Skepticism has rarely been believed by working philosophers in the strong sense. Its real role is as a stress test: any positive theory of knowledge must explain why the skeptical arguments do not refute it. The modern responses split between contextualism (the standard for knowledge varies with context, so the skeptic's exotic context does not bear on ordinary contexts), Mooreanism (we have stronger reasons to accept ordinary knowledge claims than the skeptic's premises), and externalist moves (justification does not require ruling out skeptical scenarios from inside the head).

Sub-fields and Live Areas

Contemporary epistemology has several active sub-fields.

Formal epistemology. Probabilistic and logical approaches to belief, evidence, and update. Uses tools from probability theory, decision theory, and modal logic. The dominant framework is Bayesian: belief is graded credence; rational update is conditionalization on evidence. Covered in detail at Bayesian Epistemology. Other formal frameworks include Dempster-Shafer theory, ranking theory, and AGM belief revision.

Social epistemology. How knowledge is produced, transmitted, and warranted in groups. Topics: testimony, expert disagreement, group belief, the structure of scientific communities. Goldman 1999 Knowledge in a Social World is a standard reference.

Virtue epistemology. Treats epistemic states as products of intellectual virtues (open-mindedness, intellectual courage, calibration, attentiveness). Sosa 2007, Greco 2010, Zagzebski 1996 are canonical sources. The view shifts evaluation from belief-tokens to belief-forming agents.

Naturalized epistemology. Quine 1969 argued epistemology should become a chapter of psychology: instead of normative analysis of how we should form beliefs, study empirically how we do form them. The position has been influential and divisive; many naturalized epistemologists hold a hybrid view.⁴

Epistemic injustice. Miranda Fricker's 2007 Epistemic Injustice opened a new field studying how social biases distort credibility judgments and the production of shared knowledge. Testimonial injustice (treating a speaker's testimony as less credible due to prejudice) and hermeneutical injustice (lacking the social interpretive resources to understand one's own experience) are the two central categories.

Epistemic logic. The application of modal logic to knowledge and belief. The operator $K_i \varphi$ ("agent $i$ knows $\varphi$") is treated as a modal operator with axioms specifying its formal behavior. The Fagin-Halpern-Moses-Vardi 1995 Reasoning About Knowledge is the standard reference, especially for multi-agent systems and AI applications.

The Classical JTB Analysis

The classical analysis of knowledge, traceable to Plato's Theaetetus and codified in modern form before 1963, defines:

An agent $S$ knows that $p$ iff:

1. $p$ is true.
2. $S$ believes that $p$.
3. $S$ is justified in believing that $p$.

Each clause does work. Without (1), $S$ could "know" false things, which is incoherent: knowledge, whatever else it is, is factive. Without (2), $S$ could know things she does not believe, which seems wrong. Without (3), lucky guesses count as knowledge: $S$ guesses $p$, $p$ is true, $S$ believes $p$, but the lucky-guess case is exactly what we want to exclude.

The analysis seems clean. For thirty centuries it was the working definition.

Edmund Gettier's three-page 1963 paper presented two cases where (1), (2), and (3) all hold but knowledge is intuitively absent. The cases are reproduced and analyzed at The Gettier Problem. The lesson: JTB is necessary but not sufficient. Some additional condition must be added to rule out Gettier-style cases.

The half-century of work since 1963 has produced many candidate fourth conditions. None has achieved consensus.

The half-century of failure has led some, notably Williamson, to argue that knowledge is unanalyzable: it cannot be decomposed into more basic conditions because it is itself the most basic epistemic concept. Others, such as Linda Zagzebski, argue analyses of the JTB-plus-fourth-condition form will always fail and the project should be abandoned.

The dispute is unresolved. What is settled: the simple JTB analysis is insufficient. The Gettier problem is not a dismissable curiosity; it is a substantive constraint on any theory of knowledge.

Common Confusions

Confusion 1: belief equals knowledge. Belief is a mental state in which an agent takes a proposition to be true. Knowledge requires more: the proposition must actually be true, the belief must be justified or otherwise appropriately related to truth, and (post-Gettier) some further condition must hold. Beliefs can be false; knowledge cannot. Many of the live mistakes in popular epistemology start with conflating the two.

Confusion 2: justified equals true. Justification and truth are independent. A belief can be justified but false (a confident, evidence-supported belief in a hypothesis later refuted) or true but unjustified (a lucky guess that happens to be right). Justification is a property of the agent's epistemic position; truth is a property of the world. The factivity of knowledge requires both.

Confusion 3: knowledge equals certainty. Most epistemic positions allow knowledge without certainty. I know the train will arrive at 8:15 in the same sense I know my address: I have strong evidence, no defeaters, and would assert each unhesitatingly. Cartesian-style certainty (impossible to be mistaken) is a much higher bar that few of our knowledge claims meet. Conflating the two leads either to skepticism (we cannot be certain, so we cannot know anything) or to over-confident knowledge attributions.

Confusion 4: epistemology equals philosophy of science. Philosophy of science is a sub-field of epistemology focused specifically on scientific knowledge: theory confirmation, scientific explanation, the demarcation problem, the structure of theories. Epistemology is broader: it includes everyday knowledge, testimony, perception, memory, mathematical knowledge, and self-knowledge, only some of which fall under philosophy of science.

Three Exercises

Exercise 1. For each scenario below, decide whether the agent (a) has a true belief, (b) has a justified belief, (c) has a justified true belief by the classical JTB definition, and (d) intuitively counts as having knowledge. Mark all that apply.

(i) Alice glances at a stopped clock that reads 3:00 PM and forms the belief that it is 3:00 PM. As it happens, the actual time is 3:00 PM (the clock has been broken since exactly 24 hours ago).

(ii) Bob has heard a weather forecast predicting rain and forms the belief that it will rain tomorrow, on the basis of the forecast. The forecast is from a generally reliable source. It does, in fact, rain tomorrow.

(iii) Carol is told by an unreliable acquaintance that the meeting starts at 3:00. She remembers the acquaintance is unreliable but forms the belief anyway. The meeting does start at 3:00.

(iv) Dan applies a mathematical procedure he learned in school to derive that $\sqrt{2}$ is irrational. The procedure he uses is actually flawed (a non-rigorous step), but the conclusion is true.

Exercise 2. Distinguish internalism and externalism about justification by giving one concrete case in which the two would yield different verdicts about whether an agent's belief is justified. Be specific about the relevant facts in each case.

Exercise 3. Consider this AI scenario. A language model is asked, "What was the population of Vancouver in 2020?" It produces the answer "631,486" with no qualification or hedging. Suppose the answer is exactly correct.

(a) Apply the classical JTB analysis to ask whether the model has knowledge of the population. Identify what counts as belief in this context.

(b) Apply Goldman-style reliabilism: does the model's training process count as a reliable belief-forming method for population statistics?

(c) Apply Williamson-style knowledge-first analysis: does the question of whether the model "knows" 631,486 collapse into the question of whether the model is in a position to assert it without epistemic violation?

This exercise has no single correct answer. Your job is to lay out the considerations each framework would raise.

Sketch of answers

Answer 1.

(i) Alice: (a) true belief, (b) some justification (a clock face is normally evidence), (c) JTB by classical definition. (d) The case is exactly the classic Russell counterexample: it is a Gettier-style case, true and justified but knowledge fails because of the lucky-coincidence structure. Most readers say no knowledge.

(ii) Bob: (a), (b), (c) all hold. (d) Intuitively counts as knowledge: forecast-based beliefs are the standard model.

(iii) Carol: (a) true, (b) not justified (she has reason to doubt the source), so (c) fails. (d) No knowledge: belief without justification is not knowledge.

(iv) Dan: (a) true, (b) Dan thinks he is justified, but the underlying procedure is flawed; this is a Gettier-like case for mathematical knowledge. (c) Whether (b) holds depends on whether justification is internal (Dan is justified by his lights) or external (the procedure is unreliable). (d) Intuitively, no knowledge.

Answer 2. A simple case. Sue and her duplicate both believe $p$ on the basis of the same internal evidence. Sue is in a normal epistemic environment (she got the evidence from a reliable newspaper). Her duplicate is in a demon-world where the same internal experiences are produced by a deceiver and the evidence is unreliable.

Internalism: Sue and her duplicate have the same justification, since their internal states are identical. Either both are justified or neither is.

Externalism (reliabilism): Sue is justified (her belief-forming process, reading the newspaper, is reliable in her actual environment). Her duplicate is not justified (the demon's deception is not a reliable belief-forming process).

The two positions diverge: internalism gives one verdict, externalism gives the opposite. This is the standard "evil demon" thought experiment used to make the dispute concrete.

Answer 3.

(a) JTB analysis applied to the model: The model produced a true assertion. Does the model believe the proposition? If "belief" requires anything mental in the strong sense (something it is like for the model to take to be true), the answer is contested or perhaps no. If "belief" is operationalized as the model's behavior under a wide range of probes (asking the same question under paraphrases, contexts, languages), the model probably does behave as though it believes 631,486 in a stable way for this question. The justification condition asks whether the model's basis for the answer is appropriate. Here the answer depends on the path: was the number retrieved from training data (likely), or hallucinated as a plausible value (also possible)? Without inspection of the model's actual computation, we cannot tell. The JTB analysis exposes the gap.

(b) Reliabilism: Is the trained model a reliable belief-forming process for population statistics? The empirical question. Modern LLMs are reliable for population statistics of major cities (high-frequency data in training corpora) and unreliable for population statistics of small towns or for years close to the training cutoff (where data is sparse). The reliabilist answer for this specific question (Vancouver, 2020) is plausibly yes, with the caveat that the model's reliability is not uniform across the question's instances.

(c) Williamson knowledge-first: The question becomes whether the model is in a position to assert "631,486" without epistemic violation. The standard for assertion (Williamson's knowledge norm of assertion) requires the asserter to know what is asserted. Applied to the model, this collapses the question of whether the model knows into whether the model is licensed to assert. The latter is the practically actionable question and the one product teams actually care about: under what conditions should the model be allowed to make assertions without hedging?

The exercise illustrates that traditional JTB, externalist reliabilism, and Williamson knowledge-first frame the same concrete case differently. None resolves it; together they make the dispute precise. The synthesis is at Knowledge, Justification, and LLMs.

Where Epistemology Lives in Practice

Three concrete uses.

Evidence-based medicine. EBM is in part an epistemic project: it asks what kinds of evidence (randomized controlled trials, observational studies, expert opinion) warrant clinical belief, and how the kinds should be weighted. Hierarchies of evidence (Cochrane, GRADE) operationalize epistemic standards for clinical practice. The disagreements (Should observational evidence ever beat RCT evidence? How to weigh mechanistic reasoning?) are epistemic disagreements at the highest level of practical stakes.

Statistics and AI calibration. A model's calibration (how its stated confidence relates to its empirical accuracy) is the epistemic property of greatest interest for downstream use. A model with 90 percent stated confidence on a class of cases, and 60 percent empirical accuracy on those cases, is miscalibrated in a way that misleads downstream users into treating the model as more knowledgeable than it is. This is an applied epistemology problem, and it is an active research area in modern AI.

Public-policy expert disagreement. When experts disagree (epidemiologists during a pandemic, climate scientists on regional projections, economists on policy impact), the public has an epistemic problem: how should they weight conflicting expert testimony? Social epistemology has produced formal models (Lehrer-Wagner, Hartmann-Sprenger) and substantive analyses (Goldman 2001 on testimony) that bear directly on this. The practical challenge is operationalizing them in public communication.

Prerequisites and Next Pages

• Prerequisites: What Is Philosophy?, What Is Logic?, Validity vs Soundness.
• Next: The Gettier Problem, the central technical challenge to the classical analysis of knowledge.
• Next: Induction and Hume's Problem, the foundational challenge to the rationality of empirical inference.
• Next: Bayesian Epistemology, the dominant formal framework for belief, evidence, and rational update.
• Forward synthesis: Knowledge, Justification, and LLMs, where the frameworks above are applied to current AI systems.

References

Primary texts:

• Plato. Theaetetus and Meno. Standard editions. The original sources for knowledge as more than true belief.
• Descartes, René. Meditations on First Philosophy. 1641. The foundational Cartesian internalist framework, including the canonical skeptical arguments.
• Hume, David. An Enquiry Concerning Human Understanding. 1748. Empiricist epistemology and the foundational challenge to inductive reasoning.
• Kant, Immanuel. Critique of Pure Reason. 1781, second edition 1787. The synthesis of rationalist and empiricist epistemology.
• Russell, Bertrand. The Problems of Philosophy. Williams & Norgate, 1912. Still the cleanest short introduction in English.

Modern reference:

• Gettier, Edmund L. "Is Justified True Belief Knowledge?" Analysis 23, no. 6 (1963): 121-123. The paper that broke the classical analysis.
• Goldman, Alvin I. "What Is Justified Belief?" In Justification and Knowledge, ed. George Pappas, Reidel, 1979. The founding paper of process reliabilism.
• Williamson, Timothy. Knowledge and Its Limits. Oxford, 2000. The knowledge-first framework.
• Sosa, Ernest. A Virtue Epistemology: Apt Belief and Reflective Knowledge. Oxford, 2007. The standard modern virtue epistemology.
• Audi, Robert. Epistemology: A Contemporary Introduction. Routledge, 3rd ed., 2010. Standard textbook with full coverage of the contemporary positions.
• Bonjour, Laurence. The Structure of Empirical Knowledge. Harvard, 1985. The major modern foundationalist statement.
• Quine, W. V. O. "Epistemology Naturalized." 1969. The naturalized turn.
• Fricker, Miranda. Epistemic Injustice: Power and the Ethics of Knowing. Oxford, 2007. The founding text of epistemic injustice.

Stanford Encyclopedia entries (link, do not paraphrase):

• "Epistemology."
• "The Analysis of Knowledge."
• "Internalist vs Externalist Conceptions of Epistemic Justification."
• "Reliabilist Epistemology."
• "Foundationalist Theories of Epistemic Justification."
• "Coherentist Theories of Epistemic Justification."