Epistemology

Definition and scope

Epistemology is the branch of philosophy that studies knowledge: what distinguishes knowing from merely believing, where knowledge comes from, how it is justified, and where its limits lie. The word is nineteenth-century (from the Greek episteme, knowledge, and logos, account), but the questions are among the oldest in philosophy, and they are not idle for anyone whose work consists of making claims. A physician who writes "consistent with metastatic disease" is claiming to know something about a patient. A radiochemist who reports a radiochemical purity is claiming to know something about a preparation. An informatician who records that two codes mean the same thing is claiming to know something about meaning. In each case it is fair to ask what entitles them to say so.

This monograph develops five linked ideas. Knowledge is the thing to be characterized. Belief is the attitude of holding something true. Justification is whatever turns a true belief into knowledge, if anything does. Evidence is the material on which justification draws. Uncertainty is what remains when justification is partial, which in empirical matters is always. The scope stops short of what there is (Ontology) and of how a publication records where its claims came from (Provenance and evidence).

Key distinctions

Knowledge and true belief. A guess that turns out right is not knowledge. The intuition behind every theory of knowledge is that knowing requires some connection between the believer and the fact beyond coincidence.

Justification and truth. A belief can be well justified and false. A reader who calls a lesion malignant on every available sign, and is later proved wrong by histology, was justified and mistaken. Justification is about the believer's standing; truth is about the world. Confusing them produces two errors: treating a confident conclusion as correct, and treating a correct conclusion as warranted.

Evidence and data. Data are records. Evidence is data brought into relation with a claim. The same measurement is evidence for one hypothesis, irrelevant to a second, and evidence against a third. A publication that stores claims must therefore store, for each claim, what the evidence was evidence for.

Uncertainty and ignorance. Ignorance is not knowing. Uncertainty is knowing imprecisely, and knowing how imprecisely. A measurement with a stated uncertainty conveys more than one without, not less. The Guide to the Expression of Uncertainty in Measurement makes this a discipline for physical quantities; the challenge in clinical interpretation is to do the same for judgments.

Confidence and justification. A number expressing how strongly a person or a model holds a conclusion is a report of a state, not an account of why the state is appropriate. This is the review gate for the monograph.

Historical development

Plato: knowledge as more than true belief

In the Meno, Socrates observes that a man who knows the road to Larissa and a man who merely has a true opinion about it will both get you there, and asks what the difference is worth. His answer is that true opinions run away unless tied down by an account of the reason why. In the Theaetetus, three definitions of knowledge are examined and rejected: knowledge as perception, as true judgment, and as true judgment with an account. The dialogue ends without a positive answer, but the third definition, read as "justified true belief," became the default for two millennia because the first two had been shown not to do.

Descartes: the demand for certainty

Descartes' Meditations on First Philosophy (1641) reframes the question as a search for a foundation that cannot be doubted. He subjects every belief to the possibility of error and finds that the one thing he cannot doubt while doubting is that he is thinking. Whether his rebuilding succeeds is disputed, but the method left a legacy: knowledge was to be modeled on certainty, and anything short of it was suspect. Clinical medicine, which never has certainty, has spent much of its history working out why that standard is wrong for it.

Hume: the problem of induction

Hume's Enquiry concerning Human Understanding (1748) asks a question that still has no fully satisfying answer. Every inference from observed cases to unobserved ones, from a hundred patients with a finding to the next patient with it, rests on an assumption that nature is uniform. That assumption cannot be established by reason (its denial is not a contradiction) or by experience (that would be circular). Hume concluded that induction is a habit, not a rational insight. Every statistical method is, in one way or another, an attempt to say how much the habit is worth.

Gettier: the collapse of the standard account

In 1963 Edmund Gettier published a three-page paper that ended the two-thousand-year reign of justified true belief. He gave two cases in which a person has a belief that is true, and justified, and yet is plainly not knowledge, because the justification and the truth are connected only by luck. Smith has strong evidence that Jones will get the job and that Jones has ten coins in his pocket. He infers that the man who will get the job has ten coins in his pocket. In fact Smith himself gets the job, and, unknown to him, Smith also has ten coins in his pocket. The belief is true. It is justified. It is not knowledge.

Clinical readers will recognize the pattern. A reader calls a lymph node metastatic because of apparent tracer uptake. The uptake is actually a misregistration artifact from an adjacent structure. The node, on histology, contains metastasis. The report was true, justified, and right by accident. A quality program that counts only agreement with histology will score it a success. A good department will want to know why it was right.

After Gettier: reliabilism, evidentialism, and knowledge first

The responses to Gettier define contemporary epistemology. Goldman's causal theory (1967) required that the fact be appropriately causally connected to the belief; the misregistration case fails because the metastasis did not cause the uptake. His later reliabilism (1979) relocated justification from the believer's reasons to the process that produced the belief: a belief is justified if produced by a process that reliably yields truths. A reader following a validated protocol is justified in a way a novice guessing is not, whether or not either can articulate reasons. Nozick's tracking account (1981) asked whether the belief would still be held if the fact were false; a belief that tracks the truth is knowledge. Conee and Feldman's evidentialism (1985) insisted, against reliabilism, that justification is a matter of fit between a belief and the evidence the believer possesses, whatever the process. Williamson (2000) proposed taking knowledge itself as basic rather than analyzing it into components.

None of these has won, and each captures something a working scientist recognizes. Reliabilism explains why we trust calibrated instruments and credentialed readers. Evidentialism explains why we demand that a conclusion be supported by what is actually in the record. The tracking condition explains why we value a method that would have given a different answer had the patient been different.

Probability and evidence

A parallel tradition treats belief as coming in degrees. Ramsey's 1926 essay "Truth and probability" showed that coherent degrees of belief must obey the probability calculus and can be measured by the bets a person will accept. On this Bayesian view, evidence shifts a prior degree of belief to a posterior one. Popper, in the Logic of Scientific Discovery, rejected the apparatus: hypotheses are never confirmed, only corroborated by surviving attempts at refutation. Earman's 1992 study finds the Bayesian program powerful but incomplete, especially on where priors come from.

For clinical practice the Bayesian framing has been productive, because it is the natural language of diagnostic testing: a pretest probability, a likelihood ratio, a posttest probability. Evidence-based medicine, as Sackett and colleagues defined it in 1996, asked that clinical decisions integrate the best available external evidence with individual expertise and patient values, and the GRADE framework (2008) formalized the grading of that evidence by study design, risk of bias, consistency, directness, and precision. These are applied epistemology: institutional answers to what warrants a clinical claim.

Philosophical or technical account

The structure of a warrant

Put the historical answers together and a warrant has several separable parts, each of which can fail independently.

1. The claim must be stated precisely enough that it can be true or false. "Abnormal uptake" is a claim; "uptake" is not.
2. There must be evidence: data placed in relation to the claim, with the relation stated.
3. The process that produced the evidence must be reliable for the kind of claim being made: a scanner calibrated to a standard, a reader trained on cases of the relevant kind, an assay validated in the relevant matrix.
4. The inference from evidence to claim must be valid under stated assumptions, and the assumptions must hold in the case at hand. A likelihood ratio measured in a referral population does not transfer unchanged to a screening population.
5. The connection between evidence and truth must not be accidental. This is the Gettier condition, and it is the hardest to check, because it asks a counterfactual: would the evidence have been different had the fact been different?
6. The remaining uncertainty must be characterized. A warrant that cannot say how it might be wrong is not complete.

A recorded claim that carries all six components is well warranted. Most carry one or two. The gap is the ordinary condition of knowledge in practice, but it should be visible.

Degrees and kinds of uncertainty

Uncertainty is not one thing. Measurement uncertainty, as the GUM defines it, characterizes the dispersion of values that could reasonably be attributed to a measurand. Sampling uncertainty concerns whether the cases observed represent the cases of interest. Model uncertainty concerns whether the assumptions connecting evidence to claim hold. Interpretive uncertainty concerns whether two competent readers would say the same thing. Collapsing these into one number discards the information about which kind of doubt is in play and therefore about what further evidence would reduce it.

What a confidence score is and is not

A classifier that outputs 0.92 for "malignant" has reported a number. What it means depends entirely on how it was produced. If the classifier is calibrated, the number approximates the frequency with which cases scored 0.92 were malignant in the calibration set. That is a reliabilist warrant of a specific and limited kind: the process is reliable to a stated degree in a stated population. It says nothing about whether the features that drove the score are the features that make lesions malignant, which is the causal and tracking question, nothing about whether the present case belongs to the calibration population, and nothing a reader can inspect as a reason.

A human reader's statement of confidence has the opposite profile. It may be poorly calibrated, but it can be accompanied by reasons that another reader can examine, dispute, and learn from. A justification can be given, checked, and found wanting. A score can be measured against outcomes. Both are valuable. Neither is the other. A system that records a score where a justification belongs has made a category error, and the error propagates to every downstream user who reads the number as a reason.

Biomedical relevance

Medicine is an epistemic practice before it is a therapeutic one. Every intervention rests on a chain of claims: that this finding is present, that it indicates this condition, that this condition responds to this treatment, that this patient is like the patients in whom the response was shown. Each link has a warrant with the structure above, and each has a characteristic failure mode: artifact, transfer from an unlike population, wide uncertainty reported as a point, resemblance judged on surface features. Evidence-based medicine and GRADE make the warrants for the third and fourth links explicit and gradable. They have been less concerned with the first two, the territory of diagnostic reasoning and measurement science, where nuclear medicine does most of its epistemic work.

The same analysis applies to terminology. "Code X in system A is equivalent to code Y in system B" is a knowledge claim. Its warrant requires evidence (the definitions and scope notes of both codes), a reliable process (a mapping method with documented error rates), valid inference (the definitions read in the editions actually in use), and a characterization of uncertainty (equivalence may hold for some uses and not others). A mapping asserted without these is a belief, possibly true, not knowledge. Semantic interoperability develops the consequences.

Nuclear medicine relevance

Everything in this section is a synthetic example. No real patient, product, or code is described.

A reader sits before a fluorodeoxyglucose PET/CT study of a synthetic patient restaged after treatment for an unnamed cancer. In the left supraclavicular region there is a focus of increased tracer uptake. Consider the warrant for the claim "metastatic lymph node," component by component.

The claim. "Focus of increased uptake, maximum standardized uptake value 6.1, in a 9 mm left supraclavicular node" is a claim about the image and the measurement. "Metastatic lymph node" is a claim about the patient. They have different warrants, and the report should distinguish them. (The numbers are invented and carry no clinical meaning.)

The evidence. The uptake and the node are the data. They are evidence for metastasis in virtue of a general connection: in this clinical setting, nodal uptake above a certain level is associated with malignancy at a certain rate. That connection is itself a claim with a warrant, drawn from studies in populations that may or may not resemble this patient.

The reliability of the process. The standardized uptake value is reproducible only if the scanner is calibrated, the uptake time was within protocol, the patient's glucose was measured, and the reconstruction matched the reference studies. The EANM procedure guidelines exist to make the process reliable enough that a number means something across sites. Without that, the number is a reading of a dial, not evidence.

The validity of the inference. The threshold the reader applies was derived in a population. If this patient has a condition that raises nodal uptake for other reasons, or received treatment known to produce inflammatory uptake in the interval, the assumption connecting evidence to claim does not hold. Expertise consists substantially in knowing when the assumptions fail.

The non-accidental connection. Suppose the uptake is actually a misregistration from an adjacent muscle, and the node, on later sampling, is metastatic. The report was right; the reader did not know. A department that wants to learn from its readings must record not only the conclusion but its basis, so that the basis can be checked against the later finding independently of the conclusion.

The uncertainty. The appropriate report says that the finding is suspicious for nodal metastasis, states what alternatives remain, and says what would resolve the question. PERCIST, whatever its limitations, attempts to make response claims carry stated thresholds and measurement conditions so that their uncertainty is inspectable.

Now suppose an image analysis model has been run on the same study and returns "malignant node, confidence 0.92." A calibrated score is a reliabilist warrant: on the model's validation set, cases scored near 0.92 were malignant about that often. That is useful. But the score does not tell the reader whether this study resembles the validation set, whether the features that drove the score are features of malignancy or of this scanner's reconstruction, whether the misregistration was seen or missed, or what would change the model's mind. The reader's own justification, if it is any good, answers those questions. The score answers none. It is a datum to be weighed, not a warrant to be adopted. A report that transcribes the score as the finding has replaced a justification with a report of a state.

The same structure governs a laboratory value. A creatinine result in a synthetic patient before therapy is a claim about a concentration. Its warrant depends on a calibrated assay with a stated uncertainty, a sample handled within protocol, a reference interval appropriate to the patient, and an inference from concentration to renal function that holds only under assumptions. A number in a field is the end of a chain of warrants. A system that preserves the chain lets each be asked about. One that preserves only the number cannot.

Disagreements and limitations

Internalism versus externalism. Evidentialists hold that justification depends only on what is internally available to the believer: the reasons she has. Reliabilists hold that it depends on external facts about the process, which she may not know. If reliabilism is right, an organization can confer warrant by validating processes. If evidentialism is right, a validated process without inspectable reasons produces a result the reader should not simply adopt. This publication holds that a clinical claim needs both and that systems should record both.

Bayesianism and its critics. Bayesian epistemology gives a precise account of how evidence should move belief, but it requires prior probabilities whose source is disputed. Popperian falsificationism avoids priors but gives no positive account of why a well-corroborated hypothesis should be relied on. Clinical reasoning uses both idioms without resolving the tension.

The limits of the Gettier analysis. Some philosophers hold that the search for a fourth condition that excludes luck has failed and that knowledge should be taken as primitive (Williamson). What survives for practice is the lesson that correctness and warrant must be tracked separately, which does not depend on resolving the dispute.

What this monograph does not do. It provides no reading criteria, endorses no threshold, and makes no clinical recommendation. The PET and laboratory examples illustrate the structure of a warrant, not the content of a correct reading. The historical sketches are compressed; the primary texts should be read in the editions cited, and the characterizations of the post-Gettier literature checked against the sources before this draft advances.

Related entries

• Epistemology, the lexicon entry.
• Evidence, for the relation between data and a claim.
• Uncertainty, for the kinds of uncertainty that should not be collapsed into one number.
• Ontology, for the companion question of what there is.
• Provenance and evidence, for how a publication records where a claim came from.
• Human review of machine proposals, for the principle that a proposal with a score remains a proposal.
• Human review, Provenance, Context.

References

1. Plato. Theaetetus. Levett MJ, trans; Burnyeat M, rev. In: Cooper JM, ed. Plato: Complete Works. Hackett; 1997.
2. Plato. Meno. Grube GMA, trans. In: Cooper JM, ed. Plato: Complete Works. Hackett; 1997.
3. Descartes R. Meditations on First Philosophy. In: Cottingham J, Stoothoff R, Murdoch D, trans. The Philosophical Writings of Descartes. Vol 2. Cambridge University Press; 1984.
4. Hume D. An Enquiry concerning Human Understanding. Beauchamp TL, ed. Oxford University Press; 1999.
5. Gettier EL. Is justified true belief knowledge? Analysis. 1963;23(6):121-123.
6. Goldman AI. A causal theory of knowing. Journal of Philosophy. 1967;64(12):357-372.
7. Goldman AI. What is justified belief? In: Pappas GS, ed. Justification and Knowledge. Reidel; 1979:1-23.
8. Conee E, Feldman R. Evidentialism. Philosophical Studies. 1985;48(1):15-34.
9. Nozick R. Philosophical Explanations. Harvard University Press; 1981.
10. Williamson T. Knowledge and Its Limits. Oxford University Press; 2000.
11. Ramsey FP. Truth and probability. In: Braithwaite RB, ed. The Foundations of Mathematics and Other Logical Essays. Kegan Paul; 1931.
12. Popper KR. The Logic of Scientific Discovery. Hutchinson; 1959.
13. Earman J. Bayes or Bust? A Critical Examination of Bayesian Confirmation Theory. MIT Press; 1992.
14. Sackett DL, Rosenberg WMC, Gray JAM, Haynes RB, Richardson WS. Evidence based medicine: what it is and what it isn't. BMJ. 1996;312(7023):71-72.
15. Guyatt GH, Oxman AD, Vist GE, et al. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ. 2008;336(7650):924-926.
16. Joint Committee for Guides in Metrology. Evaluation of measurement data: Guide to the expression of uncertainty in measurement (GUM). JCGM 100:2008. https://www.bipm.org/en/committees/jc/jcgm/publications (access checked 10 October 2026).
17. Boellaard R, Delgado-Bolton R, Oyen WJG, et al. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. European Journal of Nuclear Medicine and Molecular Imaging. 2015;42(2):328-354.
18. Wahl RL, Jacene H, Kasamon Y, Lodge MA. From RECIST to PERCIST: evolving considerations for PET response criteria in solid tumors. Journal of Nuclear Medicine. 2009;50(Suppl 1):122S-150S.