Classical Conditioning

Abstract

Classical conditioning is the process by which a neutral stimulus, through repeated pairing with a biologically significant event, comes to elicit a response it did not originally produce. Discovered by Ivan Pavlov in his work on canine salivation, it became the founding paradigm of the experimental study of learning and the empirical core of early behaviorism. The modern account, however, is not the reflex-stamping mechanism Pavlov's early followers assumed: Robert Rescorla showed that animals learn the contingency between events rather than their mere contiguity, and the Rescorla-Wagner model recast conditioning as error-driven learning, in which associations change only to the extent that an outcome is surprising. This article develops the core procedure, the shift from contiguity to contingency, the Rescorla-Wagner model and the blocking effect, biological constraints on association, and the neural substrates of conditioned fear and reward.

Keywords: classical conditioning, associative learning, Rescorla-Wagner model, extinction, prediction error

Classical conditioning, also called Pavlovian conditioning, is a form of associative learning in which an organism comes to respond to a previously neutral stimulus because that stimulus has reliably signalled a biologically important event (Pavlov, 1927). In Pavlov's original preparation a dog that salivated to food came, after repeated pairings, to salivate to a stimulus such as a metronome that preceded the food. The food is the unconditioned stimulus (US), the salivation it reflexively evokes is the unconditioned response (UR), the metronome is the conditioned stimulus (CS), and the salivation it comes to evoke is the conditioned response (CR). The procedure moved learning out of introspection and into the laboratory, giving psychology an objective, quantifiable index of an association forming in real time, and it supplied the young behaviorist movement with its first rigorous demonstration that even emotional reactions could be built by experience, when Watson and Rayner conditioned a fear of a white rat in an eleven-month-old infant by pairing the animal with a loud noise (Watson & Rayner, 1920). For half a century the paradigm was read as the strengthening of a stimulus-response bond by contiguity, but a series of results from the 1960s onward overturned that reading and replaced it with an account in which the animal is an active detector of the predictive relationships among events (Rescorla, 1988). The sections below present the procedure and its phases, the evidence that contingency rather than contiguity drives learning, the error-correcting Rescorla-Wagner model and the blocking effect that motivated it, the biological limits on association, and the neural circuits that implement conditioned fear and reward.

What Classical Conditioning Is

Classical conditioning is defined by its procedure and by the four elements that procedure arranges. An unconditioned stimulus is an event that evokes a reliable response without any training, such as food evoking salivation or a shock evoking freezing; the response it evokes is the unconditioned response. A conditioned stimulus is an initially neutral event that, through pairing with the US, comes to evoke a conditioned response. The conditioned response is not always a copy of the unconditioned one: a rat that freezes to a tone paired with shock is not reproducing the startle the shock itself elicits but is expressing a state of anticipatory fear organized around the predicted event (Rescorla, 1988). This distinction matters, because it separates classical conditioning from the older reflex framework in which the CS was thought simply to inherit the US's response. Pavlov's own experiments already showed that the animal responds to what the CS means rather than merely to the CS itself, and later work established that the conditioned response is shaped by the nature of the predicted outcome, its timing, and its value. Classical conditioning is thus best characterized not as the transfer of a reflex but as the acquisition of knowledge about relationships in the world: the organism learns that one event forecasts another, and it prepares accordingly. Because the response can be measured continuously as pairings accumulate, the paradigm gives an unusually direct window on the dynamics of learning, which is why it has remained a workhorse of both behavioral and neural investigation. It is one form of associative learning, and it is routinely contrasted with operant conditioning, in which behavior is controlled by its consequences rather than by a predictive stimulus. It is equally distinct from the non-associative forms of learning, above all habituation, the progressive waning of a response to a repeated, inconsequential stimulus that entails no association between two events (Thompson & Spencer, 1966). Figure 1 lays out the four elements before and after training: the US drives its response from the outset, while the CS acquires the power to evoke a conditioned response only through pairing.

The Core Procedure and Its Phases

A conditioning experiment unfolds through a characteristic set of phases, each of which reveals something about the underlying association. During acquisition, the CS and US are repeatedly paired and the conditioned response grows, rapidly at first and then more slowly as it approaches an asymptote, tracing the negatively accelerated curve that is the empirical fingerprint of associative learning. When the CS is subsequently presented alone, without the US, the conditioned response weakens across trials, a phase called extinction. Extinction was long read as the unlearning of the original association, but the modern evidence is decisive that it is not: the response returns when time passes (spontaneous recovery), when the animal is tested outside the extinction context (renewal), or when the US is presented on its own (reinstatement), which shows that the original learning survives extinction rather than being erased (Bouton, 2004). Extinction instead installs a new, context-dependent inhibitory association that competes with the original, and the balance between them is governed by the retrieval cues present at test. This has direct clinical weight, because exposure-based therapy for anxiety is a form of extinction, and the return-of-fear phenomena explain why a treated phobia can relapse. Two further phenomena complete the basic picture. In generalization, a stimulus resembling the CS evokes a conditioned response graded by its similarity, the phenomenon developed in its own right under stimulus generalization; in discrimination, differential pairing of two similar stimuli, one reinforced and one not, sharpens the response to the predictor and withholds it from the safe cue. Together these phases show that conditioning is a dynamic, reversible, context-sensitive system rather than the permanent stamping-in of a bond.

From Contiguity to Contingency

The most important conceptual change in the modern history of conditioning was the demonstration that contiguity between CS and US is neither sufficient nor necessary for learning; what matters is contingency, the degree to which the CS changes the probability of the US. Robert Rescorla made the point with a deceptively simple design (Rescorla, 1968). He held constant the probability of a shock during a tone and varied only the probability of a shock in the tone's absence. If contiguity were the engine of learning, the number of tone-shock pairings would fix the amount of conditioning, and the background rate should not matter. It mattered decisively. When shock was just as likely without the tone as with it, the tone acquired little or no fear despite being paired with shock many times, because it carried no news about the shock; when shock was likely during the tone and rare otherwise, the tone became a strong fear signal. The animal was behaving as a statistician, comparing the probability of the US given the CS against the probability of the US given the CS's absence, and conditioning tracked the difference between these two probabilities rather than their conjunction. This result cannot be accommodated by any theory in which pairings simply accumulate strength, and it forced a reconceptualization of the CS as an informational signal whose value depends on the alternatives against which it is evaluated (Rescorla, 1988). The demonstration below lets the reader set both probabilities independently and read off the resulting contingency, showing that a CS can be paired with the US on every one of its occurrences and still predict nothing.

The Rescorla-Wagner Model and the Blocking Effect

The contingency findings demanded a mechanism, and the one that reorganized the field was the Rescorla-Wagner model, the single most influential formal theory of associative learning (Rescorla & Wagner, 1972). Its central idea is that learning is driven by prediction error: on each trial the associative strengths of all present cues are summed into a prediction of the US, and each cue's strength is adjusted in proportion to the discrepancy between the US that actually occurred and this aggregate prediction. Formally, the change in a cue's associative strength is ΔV = αβ(λ − ΣV), where λ is the asymptote the US can support, ΣV is the summed strength of all cues present on the trial, and α and β are learning-rate parameters for the CS and US. When the outcome is fully predicted, λ − ΣV is zero and no learning occurs however many times the cue is paired with the US. This error term is what makes the model powerful, because it ties learning to surprise rather than to contiguity, and it predicts a family of phenomena that defeated earlier accounts. The decisive one is blocking, discovered by Leon Kamin (Kamin, 1969). If an animal is first trained until one cue, A, fully predicts the US, and a second cue, B, is then added to form a compound AB that is paired with the same US, B acquires little or no strength. The compound already predicts the US through A, so the prediction error is near zero, and there is nothing left for B to learn. Blocking is impossible to explain by contiguity, since B is paired with the US exactly as often as a control cue that is learned normally, and it is the strongest evidence that conditioning is error-driven. The robustness of the effect has itself been scrutinized: a large coordinated series of rodent studies reported fifteen consecutive failures to obtain blocking under conditions in which it was expected, which does not overturn the phenomenon but shows that even a signature prediction of the model can be procedurally fragile and boundary-dependent (Maes et al., 2016). The model also derives overshadowing, conditioned inhibition, and the overexpectation effect from the same equation, and it supplied the formal template later imported wholesale into reinforcement learning and computational neuroscience. Its limits are real and instructive: because learning depends only on the summed prediction, the basic model cannot represent within-compound associations or explain latent inhibition or spontaneous recovery, and later models added those mechanisms without discarding the error-correction core. The demonstration below runs the model trial by trial and reproduces blocking directly, letting the reader pretrain cue A and watch cue B fail to gain strength.

Biological Constraints on Association

The behaviorist programme that grew out of Pavlov assumed the equipotentiality of stimuli: any perceptible CS could be associated with any US with roughly equal ease, and the laws of learning were general across contents. John Garcia's work on taste-aversion learning dismantled this assumption (Garcia & Koelling, 1966). Rats made ill hours after consuming a novel-tasting solution formed a strong aversion to the taste, even though the sickness followed the taste by an interval far longer than the seconds within which conditioning was thought to be possible, and even after a single pairing. Crucially, the association was selective: rats readily linked taste with internal illness and audiovisual cues with external pain (shock), but not the reverse pairings. Taste predicted nausea and lights-and-sounds predicted a cutaneous threat, as if the animal came prepared to connect certain classes of stimuli and not others. This preparedness is an evolutionary adaptation, since a foraging animal that could learn in one trial to avoid a food that had poisoned it, across the long delay characteristic of digestion, would be strongly favoured. The finding was initially resisted because it violated the general-process assumption, but it is now foundational, and it reframed conditioning as a process whose parameters are tuned by the biology of the species and the ecological relationship between the stimuli. Learning is not a blank associative slate but a set of predispositions that make some contingencies easy to acquire and others nearly impossible, and any complete theory must specify the contents over which the general error-correcting rule operates.

Neural Substrates of Conditioned Fear and Reward

Because classical conditioning can be arranged with precise control over stimuli and timing, it has become the premier behavioral assay for the neuroscience of learning, and two circuits are especially well mapped. In fear conditioning, a tone paired with a mild shock comes to elicit defensive responses, and the amygdala is the site where the convergence of CS and US information is registered and stored (LeDoux, 2000). Sensory pathways carrying the tone and the shock converge on the lateral amygdala, synaptic plasticity there encodes the association, and the central amygdala drives the behavioral and autonomic components of the fear response through its downstream projections. This circuit is conserved across mammals and underlies the translational value of fear conditioning as a model for human anxiety. In reward learning, the link to the Rescorla-Wagner model became strikingly literal. Wolfram Schultz recorded from midbrain dopamine neurons and found that they do not simply signal reward but signal a reward-prediction error: they fire to an unexpected reward, shift their response to the earliest cue that predicts the reward, and fall silent, dipping below baseline, when a predicted reward is withheld (Schultz, Dayan, & Montague, 1997). This is the λ − ΣV term of the model realized as a neural signal, and it provided a mechanistic bridge between animal conditioning, the computational theory of reinforcement learning, and the study of the brain's reward system. Table 1 summarizes the four defining elements of the paradigm as they appear in these two preparations. That an abstract learning rule proposed to explain rat salivation and freezing should be found written into the firing of identified neurons is among the clearer instances of theory guiding discovery in the study of the mind.

Table 1

The Four Elements of Classical Conditioning in Two Preparations

Note. The conditioned response is organized around the predicted outcome rather than being a copy of the unconditioned response, which is why anticipatory freezing rather than startle is the conditioned reaction in fear conditioning (Rescorla, 1988).

Worked Example

The Rescorla-Wagner rule reduces to an exact calculation, and working an acquisition series and a blocking series by hand shows why surprise, not pairing, drives learning. Set the asymptote λ to 100 units of associative strength and combine the two learning rates into a single trial-level rate αβ equal to 0.30. A single cue A begins with zero strength, V equal to 0. On trial 1 the prediction is 0, the outcome is 100, and the error is 100, so ΔV is 0.30 times 100, which is 30, and V rises to 30. On trial 2 the prediction is now 30, the error is 70, ΔV is 0.30 times 70 which is 21, and V reaches 51. Continuing, trial 3 adds 0.30 times 49, which is 14.7, for 65.7; trial 4 adds 0.30 times 34.3, which is 10.29, for 75.99; trial 5 adds 0.30 times 24.01, which is 7.203, for 83.19. The increments shrink because each trial closes part of the remaining gap, tracing the negatively accelerated acquisition curve toward the asymptote of 100. Now the blocking series. Suppose cue A has already been trained to asymptote, V of A equal to 100, and a compound AB is now reinforced with the same US. On the first compound trial the summed prediction is V of A plus V of B, which is 100 plus 0, equal to 100; the outcome is 100; the error is 0. The change to B is 0.30 times 0, which is 0, and B acquires nothing on this trial or any that follows, because A already predicts the US in full. Contrast a control animal for which A was not pretrained: there the compound begins at a summed strength of 0, the first trial delivers an error of 100, and B gains 0.30 times 100, which is 30, on the very first pairing. B is paired with the US identically in the two cases; the only difference is what the other cue already predicts, and that difference is the whole of blocking (Rescorla & Wagner, 1972).

Discussion

Classical conditioning occupies a peculiar position in psychology: it is simultaneously the oldest experimental paradigm in the field and one of the most actively theorized. Its endurance is due to a reversal of its own founding interpretation. Pavlov and the behaviorists who followed him took conditioning to be the mechanical strengthening of a reflex by contiguity, and on that reading it was a narrow phenomenon of glands and muscles. The contingency experiments and the Rescorla-Wagner model turned it into something far larger, a demonstration that even simple organisms build predictive models of their world and update those models in proportion to their errors (Rescorla, 1988). This is why the paradigm has been so generative outside its original borders. The error-correction rule is formally identical to the learning signal in temporal-difference reinforcement learning, and the discovery that dopamine neurons compute a reward-prediction error joined animal conditioning to computational neuroscience and to the study of decision making under uncertainty (Schultz, Dayan, & Montague, 1997). The extinction findings, meanwhile, gave clinical psychology its mechanistic account of exposure therapy and of why treated fears relapse, since extinction adds a new inhibitory memory rather than erasing the old excitatory one, and relapse is the recovery of the original association when the retrieval context changes (Bouton, 2004). What ties these strands together is the shift from viewing the organism as a passive receiver of stamped-in bonds to viewing it as an active detector of structure, testing the predictive relations among events and revising its expectations when they are violated. That reframing, more than any single result, is why a set of experiments on salivating dogs became a foundation of the modern science of learning.

Current Directions

The most active contemporary work on classical conditioning centres on extinction as a translational model for anxiety and its treatment. A comprehensive review of the behavioral and neurobiological mechanisms of Pavlovian and instrumental extinction has consolidated the evidence that extinction is new, context-bound inhibitory learning rather than erasure, and has mapped the prefrontal-amygdala-hippocampal circuitry that gates the balance between the original and the extinction memory (Bouton, Maren, & McNally, 2021). Building on this mechanistic base, clinical researchers have argued that exposure therapy should be optimized not for within-session fear reduction, the traditional target, but for inhibitory learning: procedures that deepen the new safety association and improve its retrieval across contexts, such as varying the exposure context, combining feared cues, and removing safety signals, predict better long-term outcomes than habituation-based protocols (Craske, Hermans, & Vervliet, 2018). A parallel methodological literature has turned a critical eye on the human paradigms themselves, questioning how validly laboratory avoidance and threat-conditioning tasks capture the clinical phenomena they are meant to model and calling for better-validated measures before mechanistic claims are carried to the clinic (Krypotos, Vervliet, & Engelhard, 2018). The through-line is that the century-old paradigm is now a precision instrument for anxiety science, and the open questions are about the conditions that determine whether an extinction memory, once formed, will be retrieved when it is needed.

Commonly Confused With

Operant Conditioning

The split is antecedent versus consequence. In classical conditioning the response is elicited by a stimulus that comes before it: the CS predicts the US, and the reaction is involuntary, like salivation or freezing. In operant conditioning the response is emitted to produce a consequence that comes after it: the behavior is voluntary, and its future frequency depends on reinforcement or punishment. Ask which event the response is tied to. Pavlov's dog salivates because the bell forecasts food, whatever the dog does; Skinner's rat presses the lever because pressing produces food. When the outcome is contingent on the animal's action, the procedure is operant, not classical.

Common Misconceptions

Conditioning is just the automatic strengthening of a reflex by pairing a bell with food.

Pairing is not enough. A CS that is paired with the US on every trial acquires no strength if the US is equally likely without it; conditioning tracks the contingency between CS and US, not their contiguity (Rescorla, 1968). The reflex-strengthening picture survives because Pavlov's demonstrations were read that way for decades, but the animal is learning a predictive relationship, not stamping in a bond (Rescorla, 1988).

Extinction erases the original learning.

It does not. The conditioned response returns with the passage of time, a change of context, or a reminder of the US, which shows the original association is intact beneath a new inhibitory one (Bouton, 2004). The belief persists because the response does disappear during extinction, but disappearance at test is not erasure in storage, and the difference is why treated phobias can relapse.

Any stimulus can be associated with any outcome equally easily.

Associability is biologically constrained. Rats connect taste with illness and audiovisual cues with pain, but resist the reverse pairings, and taste aversions form across delays of hours that ordinary conditioning cannot bridge (Garcia & Koelling, 1966). The equipotentiality assumption was a premise of early behaviorism, not a finding, and it did not survive contact with the data.

Glossary

Acquisition.

The phase in which the CS and US are paired and the conditioned response grows along a negatively accelerated curve toward an asymptote.

Blocking.

The finding that a cue already predicting the US prevents a second, redundant cue paired alongside it from acquiring associative strength, because the outcome is unsurprising.

Conditioned response (CR).

The learned reaction evoked by the conditioned stimulus after conditioning; it is organized around the predicted US rather than being a copy of the unconditioned response.

Conditioned stimulus (CS).

An initially neutral stimulus that, through pairing with the unconditioned stimulus, comes to evoke a conditioned response.

Contingency.

The degree to which the CS changes the probability of the US, measured as the difference between the US probability given the CS and given its absence; the true driver of excitatory conditioning.

Discrimination.

Learning to respond to a reinforced stimulus and withhold responding to a similar unreinforced one, achieved through differential pairing.

Extinction.

The decline of the conditioned response when the CS is presented without the US; a new inhibitory learning that competes with, rather than erases, the original association.

Generalization.

The tendency of stimuli resembling the CS to evoke a conditioned response graded by their similarity to it.

Prediction error.

The discrepancy between the US that occurred and the US predicted by the cues present; the quantity that drives associative change in the Rescorla-Wagner model.

Preparedness.

The evolved predisposition to associate certain classes of stimuli with certain outcomes more readily than others, as taste with illness and audiovisual cues with pain.

Renewal.

The return of an extinguished conditioned response when the animal is tested outside the context in which extinction took place, showing extinction is context-dependent.

Rescorla-Wagner model.

The formal theory that associative strength changes in proportion to the prediction error λ minus the summed strength of all present cues, so learning depends on surprise.

Spontaneous recovery.

The reappearance of an extinguished conditioned response after a period of rest, evidence that extinction does not destroy the original learning.

Unconditioned response (UR).

The reflexive, untrained reaction evoked by the unconditioned stimulus, such as salivation to food or freezing to shock.

Unconditioned stimulus (US).

A biologically significant event that evokes a reliable response without any training and serves as the outcome the CS comes to predict.

Key Researchers

Mark E. Bouton. Professor of Psychological Science at the University of Vermont; his work on renewal, reinstatement, and spontaneous recovery established that extinction is context-dependent new learning rather than erasure. ORCID - Google Scholar - Faculty Page

Michelle G. Craske. Professor of Psychology at the University of California, Los Angeles; she reframed exposure therapy for anxiety around the inhibitory-learning model of extinction. ORCID - Google Scholar - Faculty Page - Wikipedia

John Garcia (1917-2012). Psychologist whose taste-aversion experiments overturned the equipotentiality assumption and established biological constraints on associative learning. Wikipedia

Joseph E. LeDoux (b. 1949). Professor of Neuroscience at New York University; he mapped the amygdala circuit that underlies conditioned fear. Google Scholar - Faculty Page - Wikipedia

Ivan Pavlov (1849-1936). Russian physiologist and 1904 Nobel laureate who discovered classical conditioning in the course of his research on the physiology of digestion. Nobel Biographical - Wikipedia

Robert A. Rescorla (1940-2020). Professor at the University of Pennsylvania; he demonstrated that contingency rather than contiguity drives conditioning and co-authored the Rescorla-Wagner model. Wikipedia

Wolfram Schultz (b. 1944). Professor of Neuroscience at the University of Cambridge; he discovered that midbrain dopamine neurons encode a reward-prediction error, the neural analogue of the Rescorla-Wagner error term. ORCID - Google Scholar - Faculty Page

Allan R. Wagner (1934-2018). Professor at Yale University; he co-developed the Rescorla-Wagner model and the later SOP model of associative memory. Faculty Page - Wikipedia

Frequently Asked Questions

What is classical conditioning?
Classical conditioning is a form of associative learning in which a neutral stimulus, after repeated pairing with a biologically significant event, comes to elicit a response it did not originally produce, as when Pavlov's dogs salivated to a signal that predicted food (Pavlov, 1927).

What is the difference between the CS, US, CR, and UR?
The unconditioned stimulus evokes a reflexive unconditioned response without training, such as food evoking salivation; the conditioned stimulus is an initially neutral cue that, through pairing with the unconditioned stimulus, comes to evoke a learned conditioned response (Rescorla, 1988).

Why is contingency more important than contiguity?
Rescorla showed that a stimulus paired with an outcome on every trial acquires no strength if the outcome is equally likely in its absence, so what the animal learns is the predictive relationship between events rather than their simple co-occurrence (Rescorla, 1968).

What is the Rescorla-Wagner model?
The Rescorla-Wagner model holds that an association changes in proportion to the prediction error, the gap between the outcome that occurred and the outcome predicted by all cues present, so a fully predicted event supports no further learning (Rescorla & Wagner, 1972).

What is the blocking effect?
Blocking is the finding that a cue which already fully predicts an outcome prevents a second cue, paired alongside it, from being learned, because the outcome is unsurprising and the prediction error is near zero (Kamin, 1969).

Does extinction erase a conditioned response?
No. The response returns after a delay, in a new context, or after a reminder of the outcome, which shows extinction installs a new inhibitory memory that competes with the original association rather than deleting it (Bouton, 2004).

Can any stimulus be conditioned to any outcome?
No. Associability is biologically constrained: rats link taste with illness and audiovisual cues with pain far more readily than the reverse, and taste aversions form across long delays, so learning is tuned by the species and the ecology (Garcia & Koelling, 1966).

How is classical conditioning represented in the brain?
Conditioned fear depends on the amygdala, where sensory pathways for the cue and the outcome converge, and reward learning is supported by midbrain dopamine neurons that signal a reward-prediction error matching the model's error term (Schultz, Dayan, & Montague, 1997).

References