Psychological Extinction

Abstract

Extinction is the decline of a learned response when the signal that once predicted an outcome is presented repeatedly without it, a core phenomenon of associative learning that MeSH classifies under reinforcement. Discovered by Pavlov, extinction was long read as the erasure of the original association, but decades of work show the opposite: the first memory survives, and extinction builds a second, inhibitory memory that competes with it. This is why extinguished responses come back — through spontaneous recovery over time, renewal in a new context, and reinstatement after the outcome recurs. The same principles ground exposure therapy for anxiety, reframed as inhibitory learning rather than unlearning. This article surveys the phenomenon, the new-learning account, its neural basis, and its clinical translation, with three interactive demonstrations.

Keywords: extinction, spontaneous recovery, renewal, inhibitory learning, exposure therapy

When a conditioned stimulus that has reliably predicted food, shock, or any other consequence is presented again and again with nothing following, the response it commands fades. Pavlov named this decline extinction and took it, at first, as evidence that the original learning had been undone. The century of research since has overturned that reading. Extinction turns out to be one of the most theoretically loaded phenomena in the study of learning, because the response it suppresses does not stay suppressed: it recovers with the passage of time, returns when the animal is moved to a different place, and springs back if the outcome is encountered again (Bouton, 1993). Whatever extinction does, it does not delete.

What Extinction Is

Extinction is the procedure of presenting a conditioned stimulus without its unconditioned stimulus, and the resulting decline in the conditioned response. In classical conditioning a tone that has been paired with shock comes to evoke freezing; when the tone is then sounded many times with no shock, freezing wanes. In operant conditioning a lever-press that has produced food declines once pressing no longer pays, often after a brief burst of vigorous responding first (Bouton, 2004). MeSH files the descriptor under reinforcement, a placement that captures extinction's defining condition — the withdrawal of the reinforcer or unconditioned stimulus — even though extinction is the removal of reinforcement rather than a kind of it.

The measured signature of extinction is a falling response curve: high responding at the outset, a steep early decline, and a gradual approach toward a floor. That curve invites the simplest interpretation, that the association responsible for the response is being progressively weakened toward zero, and this is exactly what early theory assumed. The interpretation is wrong, but productively so: understanding why the curve does not mean what it appears to mean is the route into the modern account (Myers & Davis, 2002).

Extinction Is Not Unlearning

The decisive evidence that extinction is new learning rather than unlearning comes from the recovery phenomena. If extinction erased the original association, an extinguished response should be gone for good; instead it returns under a range of conditions that leave the extinction training itself untouched. Rescorla made the general case that Pavlovian conditioning is the learning of relations among events, and that extinction, like acquisition, is driven by the discrepancy between what is expected and what occurs — a process that adds information rather than subtracting it (Rescorla, 1988).

On the modern view, extinction leaves the original excitatory memory intact and superimposes a second, inhibitory association: the conditioned stimulus now carries two meanings, its original one and a newer “no longer” (Bouton, 2004). Which meaning governs behavior at any moment depends on retrieval, and retrieval depends on context and time. This is why the response curve's approach to a floor is not the erasure it resembles: the excitatory memory is still there, merely out-competed for the moment by an inhibitory memory that is far more fragile and far more dependent on the circumstances in which it was learned (Bouton, 1993).

The Return of Extinguished Responding

Four phenomena together make the case that extinction spares the original memory, and each is a controlled demonstration that responding can be restored without any retraining. In spontaneous recovery, simply waiting after extinction brings the response part way back, as though the inhibitory memory decays faster than the excitatory one. In renewal, moving the animal from the extinction context to another — often the original acquisition context — reinstates responding, showing that extinction learning is tied to the place it occurred. In reinstatement, a few unsignaled presentations of the unconditioned stimulus after extinction revive the response. And in rapid reacquisition, an extinguished cue re-paired with the outcome relearns far faster than it was first acquired (Bouton, 2004).

These effects are not laboratory curiosities; in the clinic they are the return of fear that follows apparently successful treatment, and characterizing them precisely has become a methodological priority for translational work (Lonsdorf et al., 2017). The common thread is context in Bouton's broad sense — physical setting, interoceptive state, and above all time — acting as a retrieval cue that selects between the two memories the cue now evokes (Bouton, 1993).

The Neural Basis of Extinction

The recovery phenomena predict the neurobiology: if extinction is new, inhibitory learning that must be retrieved, there should be circuitry that acquires it, stores it, and gates its expression by context. Work on fear extinction in rodents and humans has identified exactly such a circuit, centered on the amygdala, the medial prefrontal cortex, and the hippocampus. The amygdala supports both the original fear association and its inhibition; the ventromedial prefrontal cortex is critical for consolidating and expressing extinction memory; and the hippocampus supplies the contextual information that decides which memory the prefrontal cortex will retrieve (Quirk & Mueller, 2008).

This division of labor makes functional sense of the behavior. Because expression of extinction depends on the ventromedial prefrontal cortex signaling safety, and because that signal is conditional on the hippocampal context, moving to a new context removes the retrieval support for extinction and renewal follows. A decade of translational research built this circuit account into a bridge between rodent lesion and human neuroimaging studies, making fear extinction one of the clearest examples of a behavioral process mapped onto a conserved neural system (Milad & Quirk, 2012). The same mechanisms operate, with parallels and differences, across Pavlovian and instrumental extinction alike (Bouton, Maren, & McNally, 2021).

Clinical Translation: Exposure Therapy

Exposure therapy — the graded, repeated confrontation of a feared stimulus in the absence of the feared consequence — is extinction carried into the clinic, and the shift from an unlearning to an inhibitory-learning model has reshaped how it is understood. If therapy does not erase the fear memory but builds a competing safety memory, then relapse is not failure but the predictable return of a fear that was never deleted, and the goal of treatment becomes making the new inhibitory memory as strong, as general, and as retrievable as possible (Craske et al., 2008).

The inhibitory-learning approach yields concrete techniques that depart from traditional habituation-based exposure. Rather than waiting for anxiety to subside within a session, the therapist maximizes the mismatch between what the patient expects and what occurs — expectancy violation — because the size of that surprise drives new learning, exactly as prediction error drives extinction in the laboratory. Varying the stimuli and settings, combining feared cues, and removing safety signals all serve to make the inhibitory memory less tied to a single context and so more resistant to renewal (Craske et al., 2014). The framework has since been consolidated into a structured set of clinical strategies, the inhibitory-retrieval approach, that operationalizes these principles for practice (Craske et al., 2022); it draws directly on extinction as a translational model of anxiety and its treatment (Craske, Hermans, & Vervliet, 2018).

Worked Example

The Rescorla-Wagner model makes extinction's central puzzle vivid. It represents the associative strength of a cue as a value V that changes on each trial in proportion to the discrepancy between the outcome that occurs, λ, and the outcome the cue predicts: ΔV = αβ(λ − V). During extinction the outcome is absent, so λ = 0 and every update is negative. Suppose a cue leaves acquisition with V = 0.9 and the learning-rate product αβ = 0.3. The first extinction trial gives ΔV = 0.3(0 − 0.9) = −0.27, so V falls to 0.63. The second gives ΔV = 0.3(0 − 0.63) = −0.189, so V = 0.441; the third takes it to 0.309, the fourth to 0.216, the fifth to 0.151.

The strength decays geometrically, multiplied by (1 − αβ) = 0.7 each trial, so that V_n = 0.9 × 0.7ⁿ — a falling curve that approaches zero but never crosses it. This is the model's success and its telling failure at once. It reproduces the smooth extinction curve, but because it drives a single associative value downward, it has no way to represent a surviving excitatory memory, and so it cannot produce spontaneous recovery, renewal, or reinstatement: once V is near zero, the model says the learning is gone. The recovery phenomena are precisely the evidence that a single decaying value is the wrong picture, and that extinction must instead lay down a second, context-gated memory alongside the first (Bouton, 2004). The arithmetic that fits the curve is the same arithmetic that fails the return of the response.

Current Directions

Because a single decaying association cannot be the whole story, recent work has asked whether extinction might be made to do more — to update or overwrite the original memory rather than merely compete with it. One influential line proposes exploiting memory reconsolidation: retrieving a fear memory can open a labile window during which extinction training may modify the original trace itself, potentially blunting the return of fear (Dunsmoor et al., 2015). The results have proven fragile and boundary-dependent, and characterizing when such effects hold has become a research program in its own right, tightly bound up with the methodological rigor the field now demands of return-of-fear measurement (Lonsdorf et al., 2017).

A second frontier targets the retrieval circuit directly. Human neuroimaging shows that the ventromedial prefrontal cortex is engaged not only in ordinary extinction but in enhanced-extinction procedures designed to strengthen and generalize the safety memory, suggesting routes to more durable clinical outcomes (Dunsmoor et al., 2019). Across both frontiers the organizing question is the one the recovery phenomena first posed: not how to weaken a fear memory, which extinction already does, but how to make the competing safety memory win reliably outside the room in which it was learned (Craske, Hermans, & Vervliet, 2018).

Key Researchers

Ivan P. Pavlov (1849-1936). Imperial Military Medical Academy, St. Petersburg; the Nobel laureate who discovered experimental extinction in the course of his classical-conditioning research, defining it as the decline of a conditioned response when the conditioned stimulus is presented without reinforcement. Wikipedia

Robert A. Rescorla (1940-2020). University of Pennsylvania; he showed that Pavlovian conditioning and extinction are the learning of relations among events driven by prediction error, and co-developed the Rescorla-Wagner model that formalized the idea. Wikipedia

Mark E. Bouton. University of Vermont; he established the context-dependence of extinction and analyzed the recovery phenomena — renewal, spontaneous recovery, and reinstatement — that show the original memory survives extinction. ORCID

Michelle G. Craske. University of California, Los Angeles; she translated extinction theory into the inhibitory-learning approach to exposure therapy, reframing the treatment of anxiety around expectancy violation and the retrievability of safety learning. Wikidata

Stephen Maren. Texas A&M University; he mapped the amygdala, prefrontal, and hippocampal circuits that support fear extinction and its context-dependent retrieval, linking the behavioral phenomena to their neural substrates. ORCID

Joseph E. Dunsmoor. University of Texas at Austin; he has advanced human extinction research, including reconsolidation-based and enhanced-extinction paradigms that test whether the original fear memory can be updated rather than merely suppressed. ORCID

Discussion

Extinction has kept its place at the center of learning theory because it is the phenomenon at which a simple, appealing idea breaks down. The idea that learning is the strengthening of an association and extinction its weakening fits the acquisition and extinction curves beautifully, and fails completely at the return of the extinguished response (Rescorla, 1988). What replaced it — that extinction is new, inhibitory learning, retrieved conditionally on context and time, competing with an original memory that never leaves — accounts for spontaneous recovery, renewal, reinstatement, and rapid reacquisition with a single principle, and it predicts the neural circuit that turns out to implement it (Bouton, 2004; Milad & Quirk, 2012).

The payoff is not only theoretical. Because exposure therapy is extinction, the inhibitory-learning model changes what a clinician is trying to do: not to delete a fear but to build a competing safety memory strong and general enough to be retrieved outside the consulting room, and to treat relapse as the expected return of a memory that was never erased (Craske et al., 2014). The open problems — whether reconsolidation can be harnessed to modify the original trace, how to make safety learning generalize across contexts, and how to measure the return of fear rigorously — are all, at bottom, versions of the question extinction posed from the start: what exactly is preserved when a learned response disappears (Bouton, Maren, & McNally, 2021)?

Glossary

Classical conditioning.

Learning in which a neutral stimulus, paired with a biologically significant one, comes to evoke a response; the setting in which Pavlov first observed extinction.

Conditioned response.

The learned reaction a conditioned stimulus comes to evoke; its decline under nonreinforcement is the observable measure of extinction.

Conditioned stimulus.

A once-neutral cue that, through pairing with an outcome, comes to predict it and to evoke a conditioned response.

Context.

In Bouton's sense, the physical setting, internal state, and time that serve as retrieval cues selecting between the excitatory and inhibitory memories a cue evokes.

Expectancy violation.

The mismatch between the outcome a patient expects and what occurs; maximizing it is held to drive inhibitory learning in exposure therapy.

Exposure therapy.

A treatment for anxiety in which a feared stimulus is confronted without the feared consequence; extinction applied in the clinic.

Extinction burst.

The temporary increase in an operant response that often precedes its decline when reinforcement is first withdrawn.

Extinction learning.

The decline of a conditioned response when the conditioned stimulus is presented repeatedly without the outcome that once followed it; the process this article treats throughout.

Inhibitory learning.

The new association formed during extinction, which signals that the cue no longer predicts the outcome and competes with the original memory.

Operant conditioning.

Learning in which behavior is shaped by its consequences; here the setting in which withdrawing reinforcement extinguishes a response.

Prediction error.

The discrepancy between the outcome expected and the outcome received; the quantity that drives both acquisition and extinction in associative models.

Rapid reacquisition.

The faster-than-original relearning of an extinguished cue when it is re-paired with the outcome, evidence the first memory survived.

Reinstatement.

The return of an extinguished response after unsignaled presentations of the unconditioned stimulus following extinction.

Renewal.

The return of an extinguished response when the animal is tested in a context different from the one in which extinction occurred.

Rescorla-Wagner model.

An associative model in which cue strength changes with prediction error; it fits the extinction curve but cannot produce the recovery phenomena.

Spontaneous recovery.

The partial return of an extinguished response after a period of rest, attributed to faster decay of the inhibitory memory.

Unconditioned stimulus.

A stimulus that evokes a response without prior learning; its withdrawal is the defining condition of extinction.

Ventromedial prefrontal cortex.

A prefrontal region critical for consolidating and expressing extinction memory, signaling safety conditional on the hippocampal context.

Frequently Asked Questions

What is extinction in psychology?
Extinction is the decline of a learned response that occurs when the conditioned stimulus, or the operant behavior, is repeatedly presented or performed without the outcome that once followed it (Bouton, 2004).

Does extinction erase the original learning?
No. The recovery phenomena show the original association survives; extinction adds a new, inhibitory memory that competes with the first rather than deleting it (Rescorla, 1988).

Why does an extinguished response come back?
Because the original excitatory memory is intact and the inhibitory extinction memory is fragile and context-bound, responding returns with time (spontaneous recovery), a change of context (renewal), or a recurrence of the outcome (reinstatement) (Bouton, 1993).

What is the difference between extinction and forgetting?
Forgetting is loss of a memory over time through disuse; extinction is active new learning produced by experiencing the cue without its outcome, and it can occur even as the original memory remains fully available (Myers & Davis, 2002).

What brain areas support extinction?
Fear extinction depends on a circuit of the amygdala, the ventromedial prefrontal cortex, and the hippocampus, with the prefrontal cortex expressing extinction and the hippocampus supplying the context that gates it (Quirk & Mueller, 2008).

How is extinction related to exposure therapy?
Exposure therapy is extinction applied clinically: confronting a feared stimulus without the feared consequence builds a competing safety memory, which is why the inhibitory-learning model now guides its practice (Craske et al., 2008).

What is the inhibitory-learning approach to exposure?
It optimizes exposure by maximizing the violation of the patient's expectations and by varying stimuli and settings, so the new safety memory is strong and retrievable across contexts rather than tied to one (Craske et al., 2014).

Can the original fear memory ever be modified?
Possibly. Reconsolidation-based and enhanced-extinction procedures aim to update the original trace during a labile retrieval window, but the effects are fragile and their boundary conditions remain under active investigation (Dunsmoor et al., 2015).

References