Evidence is whatever provides a reason to accept, reject, or revise a claim. It includes observations, experiences, measurements, documents, testimony, and physical traces, although philosophers disagree about which of these are evidence in the strictest sense. A fingerprint may be evidence that someone touched an object; a change in a spectrum may be evidence about a star’s composition; and a remembered conversation may be evidence about a promise. Evidence is relational: something is evidence for or against a proposition, usually against a background of assumptions and competing explanations.
The concept is central to epistemology, where it connects belief with knowledge and rational justification. It also organizes scientific investigation, historical reconstruction, and legal fact-finding. These settings share a concern with warranted conclusions but differ in their purposes, procedures, and standards. Evidence sufficient to pursue a scientific hypothesis may be insufficient to establish criminal guilt, while information excluded from a courtroom may remain relevant to a historian.
Understanding evidence therefore requires distinguishing what evidence consists of, how it supports conclusions, who possesses it, and what makes it trustworthy. It also requires separating evidential support from certainty: evidence can be genuine yet misleading, strong yet incomplete, or adequate for one question without resolving another.
The nature and structure of evidence
In ordinary usage, “evidence” can refer either to material objects or to what those objects indicate. A blood sample is called evidence, but its significance may lie in a laboratory result, such as the presence of a particular genetic sequence. A document can be an evidential object, while the proposition that it bears a certain signature is an evidential claim. Distinguishing these levels prevents confusion between preserving an object, accurately describing it, and correctly interpreting its implications.
Philosophical accounts differ over the basic category to which evidence belongs. On propositional accounts, evidence consists of propositions that can serve as premises in reasoning. On experiential accounts, perceptual experiences can themselves justify beliefs without first being converted into beliefs about those experiences. Other approaches emphasize publicly accessible objects and events, especially when explaining scientific inquiry. These accounts answer somewhat different questions: what justifies an individual’s belief, what can be shared in an argument, and what investigators can inspect.
Evidence normally has at least three components: an evidential item , a claim , and background information . A wet pavement may support the claim that it rained, but not if the background includes knowledge that a street-cleaning vehicle has just passed. Likewise, an instrument reading supports a conclusion only through assumptions about calibration, operating conditions, and the connection between the measured quantity and the phenomenon of interest. Evidential relevance is therefore rarely an intrinsic property of an isolated object.
Several distinctions concern the direction and strength of support. Positive evidence favors a claim; negative evidence counts against it. Direct evidence, especially in legal terminology, bears on a disputed fact without requiring the particular intermediate inference characteristic of circumstantial evidence. This does not make it infallible: an eyewitness can be mistaken, while a combination of circumstantial traces can be highly compelling. Evidence can also be comparative, favoring one explanation over another without making either explanation probable in absolute terms.
Evidence must further be distinguished from truth, persuasion, and proof. A statement can be true without anyone having evidence for it, and a person can be persuaded by something that provides little genuine support. In mathematics, a proof aims to establish a conclusion through valid inference from accepted premises; empirical evidence generally supports conclusions that remain revisable. Neither quantity nor vividness alone determines evidential force. A thousand copies of one inaccurate report do not provide the same support as a thousand independent observations.
Evidence, justification, and knowledge
Evidence is frequently understood through its role in epistemic justification: the status that makes a belief reasonable from an intellectual point of view. Evidentialist approaches hold, broadly, that justified belief depends on how well it fits the evidence available to the believer. On such accounts, believing too strongly on weak evidence is an epistemic defect even when the belief happens to be true. Conversely, a belief can be reasonable given the available evidence and nevertheless turn out false.
A central distinction separates having justification for a proposition from holding a belief on the right basis. Someone may possess excellent evidence that a bridge is unsafe but believe this solely because of an irrelevant superstition. The proposition is supported by evidence available to that person, yet the actual belief does not appropriately depend on it. Philosophers call this the distinction between propositional and doxastic justification. It explains why merely placing evidence in a person’s possession does not automatically make every corresponding belief well-founded.
Disagreement also concerns what counts as possessing evidence. Must a person consciously attend to it? Can forgotten but retrievable information count? Does an unnoticed implication of known facts belong to the person’s evidence? These questions become important when assessing complex reasoning. A researcher may possess a dataset without recognizing an important pattern, and a witness may have seen an event without retaining an accurate memory. Practical assessments often distinguish information available in principle from information that a reasoner can actually use.
Foundationalism and coherentism offer different accounts of how evidential support is organized. Foundationalist theories allow some beliefs to receive justification without depending on further justified beliefs, often through perception or experience. Coherentist theories emphasize support within an interconnected system. Reliabilism shifts attention toward whether beliefs arise through processes that reliably produce truth, rather than exclusively toward the evidence accessible to the believer. These approaches can disagree about a person whose cognitive process is dependable but whose reasons are difficult to articulate.
The relation between evidence and knowledge is especially contested. In Knowledge and Its Limits (2000), Timothy Williamson defended the thesis that a person’s evidence consists of what that person knows, commonly expressed as . This excludes false propositions from evidence in the strict sense. Other theories allow misleading experiences or justified false beliefs to function as evidence. The Gettier problem, introduced in Edmund Gettier’s 1963 paper, meanwhile shows that a true belief can be justified yet true through epistemic luck. Evidence must therefore be connected appropriately to truth if it is to explain knowledge, rather than merely rational confidence.
Historical development of the concept
The English word “evidence” derives through French from Latin evidentia, associated with visibility or manifestness. Its history reflects a persistent tension between what appears self-evident and what must be established through signs, testimony, or inference. Ancient rhetoric treated persuasion in courts and assemblies as a central problem, while ancient philosophy examined the difference between knowledge, opinion, and demonstration. These traditions did not employ a single concept identical to every modern use of “evidence.”
In Aristotle’s Posterior Analytics, scientific understanding involves demonstration from premises that are true, primary, and explanatory. His Rhetoric, by contrast, discusses forms of persuasion suited to practical disputes, including signs, examples, and witness testimony. This distinction anticipates a durable difference between demonstrating why something must be so and assembling reasons for a conclusion under uncertain conditions. Ancient discussions also recognized that signs vary in reliability: an observed effect need not uniquely identify its cause.
Early modern debates transformed the relation between evidence and experience. René Descartes sought certainty through clear and distinct perception, especially in the Meditations on First Philosophy (1641). John Locke, in An Essay Concerning Human Understanding (1690), distinguished knowledge from probable judgment and examined how assent should reflect the grounds available. David Hume analyzed causal inference and testimony in An Enquiry concerning Human Understanding (1748). His treatment of miracles made the comparative credibility of testimony and established experience a prominent problem.
The development of probability supplied another framework. Thomas Bayes’s essay, published posthumously in 1763, addressed an inverse problem of probability; Pierre-Simon Laplace subsequently developed methods for reasoning from observed effects to uncertain causes. These developments did not immediately produce a unified theory of evidence, but they helped make uncertain inference mathematically tractable. During the nineteenth century, debates about testimony, induction, and scientific method increasingly interacted with statistical reasoning.
Twentieth-century philosophy shifted attention toward confirmation, testing, and the relation between observation and theory. Logical empiricism sought systematic accounts of empirical meaning and support. Rudolf Carnap developed formal approaches to inductive logic, while Karl Popper emphasized attempts to refute theories rather than verify them. Thomas Kuhn highlighted the historical role of disciplinary frameworks in interpreting observations. These approaches left enduring questions: whether evidential support can be captured by a universal calculus, whether observation is independent of theory, and how rational assessment changes during major scientific transitions.
Forms of evidential inference
Evidence supports conclusions through different kinds of reasoning. In deductive reasoning, the conclusion follows necessarily if the premises are true and the inference is valid. A laboratory report and a classification rule might jointly entail that a sample meets a specified criterion. But deduction alone does not establish the truth of the report or the appropriateness of the classification rule. Logical validity concerns the preservation of truth from premises to conclusion, not the independent credibility of the premises.
Inductive reasoning extends beyond what observations strictly entail. Repeated observations of a material expanding when heated may support a generalization about that material. Sampling allows conclusions about a population from observations of a subset, provided the sampling process is appropriate. The problem of induction concerns how such extensions can be justified without assuming, circularly, that previously successful patterns of inference will continue to work. Actual scientific induction is constrained by background theory, measurement quality, and the selection of cases, rather than merely by the number of observations.
Abductive reasoning, often called inference to the best explanation, evaluates hypotheses by their ability to explain the evidence. A geologist may compare several mechanisms capable of producing a rock formation. Explanatory breadth, consistency with established findings, simplicity, and predictive success can influence the comparison. Yet an explanation’s attractiveness does not guarantee its truth: the best available explanation may be the best of an inadequate set. Invoking Occam’s razor therefore requires care, because fewer assumptions do not automatically imply greater empirical accuracy.
Causal inference presents a particularly important case. An association between two variables can arise because one causes the other, because a third variable influences both, or because selection procedures create an apparent relationship. Evidence for causation must discriminate among these possibilities. Experiments can intervene on a proposed cause, while observational studies may use temporal structure, comparison groups, natural experiments, or explicit causal models. Each approach depends on assumptions that can themselves require evidence.
The cumulative force of evidence depends on its structure. Independent findings can provide substantial mutual reinforcement; dependent findings may add little. A newspaper report, a social-media post quoting it, and a witness who learned the story from that post are not three independent sources. Similarly, several tests may share one faulty calibration procedure. Evaluating a body of evidence therefore requires tracing information pathways and common causes, not simply counting favorable items.
The total-evidence principle holds that assessment should take account of all relevant available evidence rather than a selectively favorable subset. This does not require retaining every detail with equal weight. It requires explaining why some information is reliable, irrelevant, redundant, or outweighed. A conclusion supported by one striking observation may become unreasonable when placed alongside a larger and better-controlled body of contrary results.
Probability and the measurement of support
Probability provides tools for representing uncertain belief and assessing how observations discriminate among hypotheses. Bayes’s theorem states, for a hypothesis , evidence , and background information ,
provided the denominator is nonzero. The posterior probability combines the prior probability of the hypothesis with the likelihood of observing the evidence if that hypothesis is true. The formula is mathematically uncontroversial; disagreements concern its interpretation, its application, and the assignment of probabilities.
One common Bayesian criterion says that confirms when . Confirmation in this sense means an increase in probability, not proof or even high probability. Evidence can raise a hypothesis’s probability from one in a million to one in a thousand while leaving it unlikely. Conversely, a highly probable hypothesis may receive little additional confirmation from an observation already expected under many alternatives.
The likelihood ratio measures how well evidence distinguishes two hypotheses:
If the evidence is ten times as probable under as under , the likelihood ratio is 10. Posterior odds equal prior odds multiplied by this ratio. This makes clear why evidence cannot usually be evaluated by asking only whether it is compatible with a favored hypothesis. Evidence expected under every serious alternative has little discriminatory value.
Base rates illustrate the importance of prior information. Suppose a condition occurs in 1% of a population and a test has 99% sensitivity and 95% specificity. Among 10,000 people, the expected counts are 99 true-positive and 495 false-positive results. Under those assumptions, the probability that a randomly selected person with a positive result has the condition is , approximately 16.7%. The example is a mathematical illustration of how test performance and prevalence interact, not a description of any particular diagnostic procedure.
Frequentist statistics evaluates procedures through their behavior under repeated sampling. A -value is the probability, under a specified null model, of obtaining a test statistic at least as extreme as the observed one. It is not the probability that the null hypothesis is true, nor the probability that the result occurred “by chance.” Confidence intervals likewise have a repeated-sampling interpretation rather than automatically assigning a probability to a fixed parameter after the data are observed.
No single number captures every aspect of evidential quality. A small -value does not establish a large effect, sound measurement, freedom from selection bias, or an appropriate causal interpretation. Bayesian results also remain sensitive to model assumptions and, in some settings, prior choices. Formal measures clarify particular relations among assumptions, data, and conclusions; they do not eliminate the need to evaluate how the data were produced.
Evidence in scientific investigation
Scientific evidence is produced through organized interactions among observation, instruments, models, and criticism. The scientific method is not one invariant sequence, but scientific disciplines commonly require explicit procedures, checks on error, and opportunities for independent scrutiny. A telescope image, a particle-detector event, and a laboratory assay become evidentially useful through knowledge of how their instruments respond. Raw output does not interpret itself: calibration, background subtraction, uncertainty estimation, and classification can all affect what conclusions it supports.
An experiment strengthens inference by manipulating conditions and comparing outcomes. A randomized controlled trial uses random assignment to reduce systematic differences between comparison groups, thereby improving the basis for estimating causal effects. Randomization balances confounding factors in expectation, not necessarily in every realized sample. Failures of allocation concealment, missing outcomes, nonadherence, and departures from planned analysis can weaken an inference even when assignment was originally random.
Evidence also comes from phenomena that cannot be experimentally manipulated. Astronomy investigates distant objects; geology reconstructs past processes; evolutionary biology draws on fossils, genetic sequences, and present-day populations. Such fields use converging observations, distinctive predictions, and comparisons among explanatory models. Historical sciences are not merely collections of untestable stories: claims about past events can entail observable consequences, such as particular isotope patterns or distributions of fossils, that distinguish them from alternatives.
Reproducibility and replication help identify whether a result depends on computational mistakes, peculiar samples, or unrecognized conditions. Terminology varies across disciplines, but a common distinction separates reproducing an analysis with the same data and methods from obtaining consistent findings with new data. A failed replication does not automatically identify which study is mistaken. Differences may reflect random variation, weak measurement, changed conditions, or an effect that depends on context. Repeated, well-designed investigations help discriminate among these possibilities.
Scientific synthesis evaluates bodies of findings rather than treating every published result as equally informative. A systematic review uses explicit methods to identify, select, and appraise relevant studies. Meta-analysis can statistically combine results when their designs and outcomes are sufficiently comparable. Publication bias, selective reporting, and heterogeneity can distort pooled estimates. Large numbers of studies do not compensate automatically for shared methodological defects.
The 1919 eclipse observations illustrate both the importance and the complexity of a celebrated test. Expeditions compared the apparent positions of stars near the Sun with reference positions to assess gravitational light deflection associated with general relativity. The observations involved difficult measurement conditions and differing instrumental quality; later measurements provided much stronger constraints. The case shows why scientific evidential standing can evolve: a historically influential result may initiate acceptance without remaining the strongest available support for the theory.
Testimony, records, and historical knowledge
Much human knowledge depends on testimony: information acquired from what others say or record. No individual directly verifies every geographical fact, scientific result, or event in the past that they reasonably accept. Testimony is therefore not an exceptional substitute for firsthand knowledge but a basic component of social epistemology. Its evidential force depends on the speaker’s access to information, competence, sincerity, and relationship to other sources.
Philosophers disagree about whether testimonial justification must ultimately be reduced to independent evidence of a speaker’s reliability. Reductionist approaches emphasize supporting observations about speakers and communication practices. Non-reductionist approaches allow a basic entitlement to accept testimony in the absence of defeaters, while retaining sensitivity to deception and error. Both approaches must explain why trust is indispensable yet selective. A witness may report an event honestly but lack the perceptual conditions needed to identify a person accurately.
Historical inquiry combines testimony with documents, material remains, and contextual knowledge. Source criticism distinguishes questions of authenticity, authorship, date, purpose, transmission, and content. Establishing that a letter is genuine does not establish that every claim in it is true. A forged document can also provide evidence about the period and circumstances of its fabrication. The evidential value of a source is thus question-dependent: propaganda may be unreliable about a military victory while highly informative about a government’s desired public image.
Independent corroboration is particularly important. Two chronicles may appear to confirm one another but actually derive from a common source. Conversely, disagreement does not necessarily make both worthless; it may expose differences in perspective, terminology, chronology, or access. Archaeology adds material constraints through stratigraphy, dating methods, settlement patterns, and artifacts. Documentary and archaeological evidence can reinforce or challenge each other without either category receiving automatic priority.
Absence can also be evidence, but only relative to an expectation of detection. The absence of an event from a surviving archive matters more when comparable events were routinely recorded and the relevant records are substantially complete. It matters less when records were selective, destroyed, or never produced. Formally, a failure to observe counts against a hypothesis when would probably have been observed if the hypothesis were true.
Digital records introduce additional problems of provenance and independence. A timestamp may reflect upload time rather than creation time; copied images can lose metadata; apparently separate accounts can repeat a single original claim. Oral traditions likewise require attention to transmission practices, performance contexts, and changes over time rather than either blanket acceptance or blanket dismissal. Across different media, evaluating evidence requires reconstructing how information traveled from an event to the surviving record.
Evidence in law and adjudication
In law, evidence is regulated information presented to establish or contest facts relevant to a legal proceeding. Legal systems distinguish substantive questions—what conduct creates liability or constitutes an offense—from evidentiary questions concerning what information a decision-maker may consider and how it may be presented. Legal evidence therefore serves institutional purposes beyond finding truth in the abstract, including procedural fairness, protection of rights, and the orderly resolution of disputes.
Relevance is usually necessary but not sufficient for admissibility. In United States federal proceedings, Federal Rule of Evidence 401 defines relevant evidence through its tendency to make a consequential fact more or less probable. Rule 403 permits exclusion when probative value is substantially outweighed by specified dangers, including unfair prejudice, confusion, or misleading the jury. Other jurisdictions organize these questions differently. A vivid photograph, for example, may convey relevant information while also raising concerns about disproportionate emotional influence.
Rules concerning hearsay, privilege, and unlawfully obtained material demonstrate the difference between evidential relevance and legal admissibility. Hearsay rules regulate certain out-of-court statements offered for their truth and contain numerous qualifications and exceptions. Privileges can protect confidential relationships even when the information withheld would be relevant. Exclusionary rules concerning unlawful searches vary across legal systems and circumstances. These doctrines do not imply that excluded material is always false or that admitted material is always trustworthy.
Burdens and standards of proof determine who must establish a proposition and to what degree. Common-law criminal proceedings generally require the prosecution to prove guilt beyond a reasonable doubt. Many civil claims use a balance-of-probabilities or preponderance standard. Some matters require intermediate standards, such as clear and convincing evidence in specified United States proceedings. These verbal standards are not ordinarily fixed numerical probability thresholds; their legal meaning depends on jurisdiction, case law, and the issue being decided.
Physical and digital exhibits require attention to authentication and integrity. A chain of custody documents the collection, transfer, storage, and examination of an item. Breaks or uncertainties can raise questions about contamination, substitution, or alteration, although their legal consequences depend on the circumstances. DNA evidence illustrates the distinction between identifying a biological source and explaining an event. A genetic association may strongly bear on whose material was present without independently establishing when or how it arrived, or whether an offense occurred.
Expert evidence presents further questions about methods, error rates, and the limits of interpretation. In Daubert v. Merrell Dow Pharmaceuticals (1993), the United States Supreme Court interpreted the federal trial judge’s gatekeeping role concerning scientific expert testimony. Statistical evidence also risks the prosecutor’s fallacy: confusing the probability of a match under an alternative hypothesis with the probability that the alternative hypothesis is true given the match. Proper evaluation requires keeping the evidential comparison separate from the ultimate legal conclusion.
Underdetermination, disagreement, and evidential limits
Evidence can leave more than one explanation open. Underdetermination occurs when the available evidence does not uniquely determine which of several hypotheses is correct. A finite sequence of observations may fit multiple mathematical models, and the same historical traces may be compatible with different reconstructions. This does not show that every explanation is equally good. Alternatives can differ in fit with other evidence, explanatory reach, complexity, or capacity to generate new tests.
The dependence of observation on assumptions creates a related problem. An unexpected result can indicate that a central theory is false, but it can also reflect a faulty instrument, an incorrect auxiliary assumption, or an unsuitable statistical model. Confirmation holism emphasizes that testing commonly involves networks of claims rather than isolated propositions. This complicates simple accounts of falsifiability, while still allowing empirical criticism: auxiliary assumptions can be tested independently, and competing explanations can be compared across multiple settings.
Evidence can be defeated in different ways. A rebutting defeater supports the opposite conclusion: evidence that a train has been canceled rebuts the expectation that it will depart. An undercutting defeater attacks the connection between evidence and conclusion: learning that a departure display is malfunctioning undermines reliance on the display without itself proving cancellation. Higher-order evidence concerns the quality of one’s own assessment, such as learning that a calculation was performed under conditions known to increase errors. Such information can warrant reconsideration even before the original calculation is rechecked.
Persistent disagreement raises questions about whether the same evidence can rationally support different conclusions. Some theories allow a range of reasonable responses because people may differ in prior probabilities, evidential weighting, or standards of explanation. Others argue that rationality more tightly constrains belief. Disagreement among similarly competent investigators can itself be relevant evidence, but its significance depends on whether their judgments are independent and whether they genuinely share the same information.
Human evaluation is also affected by confirmation bias, the tendency to seek or interpret information in ways favorable to an existing view. Institutional structures can amplify such effects through selective publication, restricted access, or incentives to produce particular findings. Conflicts of interest do not by themselves establish that a claim is false, but they can bear on the reliability of the process producing it. Epistemic injustice concerns, among other things, credibility judgments distorted by prejudice, which can prevent relevant testimony from receiving appropriate weight.
Finally, evidential sufficiency depends partly on the question being asked. The evidence needed to distinguish two mechanisms is not necessarily the evidence needed to estimate an effect’s magnitude or decide a legal case. Decision-making additionally involves possible consequences and values, which should not be confused with the truth of the underlying claim. Fallibilism allows well-supported knowledge and rational belief without demanding immunity to every possibility of error. On this view, acknowledging uncertainty is compatible with making substantive distinctions between conjecture, credible support, and strongly established conclusions.