Research Brief · The Evidence, Plainly
Safety Has to Be Learned
When a threat response finally stops firing, the brain has not deleted the lesson that produced it. It has built a second, competing lesson beside the first one, and that newer lesson is tied to the room, the hour, and the state it was learned in. That is why an old state can return in a new context without meaning the work failed.
What was actually studied
A synthesis of the Pavlovian extinction literature, asking whether extinction destroys the original learning or generates new learning. The evidence examined includes the three recovery effects (spontaneous recovery, renewal, and reinstatement) and what they reveal about how the new learning is stored and retrieved. Largely animal literature, with direct human parallels.
More than 500 participants across a range of learning paradigms, with functional, effective, and structural connectivity measured across the core learning network (amygdala, hippocampus, anterior cingulate, ventromedial prefrontal cortex, cerebellum). The question was whether the same connections predict how well a person acquires a learned threat and how well they extinguish it.
A three-day fear acquisition and extinction protocol in healthy adults. Anodal transcranial direct current stimulation was applied over the ventromedial prefrontal cortex during the extinction learning phase (21 stimulated, 23 sham), with skin conductance and simultaneous functional MRI recorded throughout.
Transcranial ultrasound stimulation targeting the bilateral amygdala during Pavlovian threat conditioning in healthy adults (25 in the amygdala condition, 25 in a hippocampus control condition), with trial-level modeling of skin conductance to separate how fast a threat was acquired from how readily it extinguished.
What they found
Extinction is new learning, not erasure. When a cue stops predicting something bad and the threat response fades, the original association survives. The evidence is in the recovery effects. Time passes and the response comes back (spontaneous recovery). The setting changes and the response comes back (renewal). An unrelated stressor arrives and the response comes back (reinstatement). Bouton’s conclusion is that extinction generates a second, inhibitory memory that competes with the first, and that the second memory is unusually dependent on the context in which it was learned (Bouton).
That second lesson has an address. Human work has placed the retrieval of extinction memory in the ventromedial prefrontal cortex, and a 2026 experiment tested that circuit directly rather than observing it. Healthy adults received either real or sham electrical stimulation over the ventromedial prefrontal cortex while they learned that a previously threatening cue was now safe. The stimulated group showed significantly better extinction recall the next day, along with earlier engagement of safety-related regions and reduced coupling between the prefrontal cortex and the fear network (Chhabra et al.).
The two processes are also asymmetric in speed. Using focused ultrasound to modulate the amygdala during threat conditioning, a 2026 study found that quieting the amygdala slowed the initial acquisition of a threat and made the later extinction easier, and described the default emotional learning state in four words: learning fast, forgetting slow (Meijer et al.). Read alongside the connectivity findings, this reframes a familiar experience. Being quick to alarm is not the same trait as being poor at unlearning, and the two do not have to travel together.
Where the evidence stands
Four sources, one throughline.
One establishes the architecture: extinction builds a second, context-bound memory rather than deleting the first (Bouton). One shows the second memory has a circuit that can be strengthened, and that strengthening it improves recall of the safety lesson a day later (Chhabra et al.). One shows that acquiring a threat and unlearning it are predicted by different connectivity, in a sample large enough to take seriously (Gomes et al.). And one shows the asymmetry directly, with the threat side of the system built to learn fast and let go slowly (Meijer et al.).
Together they change what the return of an old state means. It is not evidence that the work failed or that the original response was never really addressed. It is the expected behavior of a system that keeps both lessons on file and retrieves whichever one fits the current context. The practical consequence is that the useful work is not erasure. It is giving the second lesson enough repetition, in enough different settings, and under conditions where the prefrontal circuit that carries it is actually online.
What this does not prove
The human experiments here are small and tightly controlled. The stimulation study ran 44 participants total across two groups in a laboratory conditioning task (Chhabra et al.), and the ultrasound study ran 25 per condition (Meijer et al.). Both use mild laboratory threats and healthy adults, which is a long way from a real threat memory formed in a real life. The large connectivity study is correlational and predictive rather than causal, and it pools across paradigms (Gomes et al.). Bouton’s synthesis rests substantially on animal work, with human replication of the recovery effects but not of every mechanism.
None of this is diagnostic, and none of it is a treatment protocol. This is education, not medical advice. Extinction is the mechanism that exposure-based therapies are built on, and running that work alone, at a dose that overwhelms rather than teaches, is a known way to make things worse rather than better. If threat responses are disrupting daily life, that work belongs with a trauma-informed clinician.
What it means for you
If an old state keeps returning, the return is information about context rather than a verdict on your progress. Three things follow. First, record the conditions rather than the self-assessment: where you were, what time it was, who was present, what state you walked in with. Renewal is context-bound, so the context is the data that can actually be worked with. Second, practice the alternative in more than one setting, because a second lesson learned in a single place tends to stay in that place. Third, protect the circuit that carries the new learning. It goes offline under acute stress, which is why regulating first and approaching second is the order that works, and it consolidates during sleep, which is why a short night trims the stage that does the filing. The newsletter lays out the five moves. The research above is why they work.
Go to the source
- Tier 2 · Review Bouton, M. E. (2004). Context and behavioral processes in extinction. Learning & Memory, 11(5), 485-494.
doi.org/10.1101/lm.78804 - Tier 2 · Imaging Gomes, C. A., Bach, D. R., Razi, A., et al. (2026). Predicting individual differences of fear and cognitive learning and extinction. Nature Communications, 17(1).
doi.org/10.1038/s41467-026-71830-0 - Tier 2 · Experiment Chhabra, H., Ma, Y., Genç, E., Nitsche, M. A., & Yavari, F. (2026). The physiological foundation of extinction improvement by tDCS over the ventromedial prefrontal cortex in healthy humans. Translational Psychiatry, 16(1).
doi.org/10.1038/s41398-026-04190-4 - Tier 2 · Experiment Meijer, S., Carpino, E., Kop, B. R., et al. (2026). The human amygdala in threat learning and extinction. Science Advances, 12(13), eaea8233.
doi.org/10.1126/sciadv.aea8233
Tier 1 means peer-reviewed primary research or meta-analysis, the strongest evidence. Tier 2 means an expert framework or smaller study that traces to peer-reviewed work. We grade every source so you can see the weight behind each claim.
Pegasus Realm publishes Wellness Team Culture and the practice resources behind it.
