Look at the research literature on fear and one thing is immediately obvious: almost all of it is about threat. How threat is detected, learned, generalised, remembered, extinguished. Safety appears mostly as the control condition — the thing that happens when the threat is not there. Reversing the question turns out to be more interesting than it sounds, because safety appears not to be an absence at all.

The Assumption

The intuitive model is a switch. Threat present, alarm on; threat absent, alarm off. On that picture, feeling safe is a default state and the interesting machinery is all on the threat side.

It is a reasonable assumption and it makes a testable prediction: if safety were merely the absence of threat, there would be nothing separate to look for. The brain would simply not be doing the thing it does when threatened.

The Test

Human neuroimaging studies of the amygdala in threat conditioning had produced inconsistent results, which is often a sign that the question is being asked at the wrong resolution. A 2022 paper identified three candidate reasons: the temporal profile of the amygdala response, the anatomical specificity of responses during threat versus safety learning, and insufficient power to detect either.

To address the power problem it combined data across multiple studies using a well-validated human threat conditioning paradigm, reaching 601 humans (PNAS, 2022).

The paradigm itself is simple. One stimulus is paired with an aversive shock; another is not. Learning the second is safety learning — and it is learning, not the absence of it.

What They Found

Two amygdala subregions tracked the shock-paired stimulus during early conditioning. But only one demonstrated delayed responding to the stimulus that was not paired with shock.

Two things follow. First, the anatomy differs — safety learning did not simply engage the same regions less. Second, the timing differs: the response to the unpaired stimulus was delayed relative to the threat response.

A delay makes sense once you consider what each computation requires. Establishing that something is dangerous can be done in one trial, and being wrong in the cautious direction is cheap. Establishing that something is reliably not dangerous requires accumulating evidence of nothing happening, repeatedly — which takes longer by necessity.

The paper frames the result as identifying cross-species similarities in temporally and anatomically specific amygdala contributions to threat and safety learning, affirming human amygdala involvement in associative learning.

Why the Asymmetry Matters

If safety takes longer to learn than threat, and uses partly different machinery, then reassurance is not the inverse of alarm and cannot be expected to work as quickly.

It also reframes what a safe environment does. On the switch model, safety is what you get by removing threats. On this account, safety is something a nervous system has to accumulate evidence for — which means repetition and predictability are not comforts layered on top of the real thing, they are the mechanism.

That connects directly to why predictability does more work than getting things right: a system learning safety is a system counting how often what was expected actually happened.

Where the Question Goes Next

A 2026 study moves from regions to networks, and from adults to youth. Its motivation is stated plainly: affective disorders rise sharply in prevalence during puberty, a period marked by heightened affective sensitivity and by ongoing development of fronto-amygdala circuitry.

Rather than asking which regions activate, it compares connectivity within and between three large-scale networks — the salience network, the executive control network and the default mode network — through fear learning, in young adults with and without mild mixed affective complaints (British Journal of Psychiatry, 2026).

The authors are explicit that they are looking for potential neural targets for approaches such as fMRI-guided neurofeedback or transcranial magnetic stimulation. That is a statement of research intent, not of available treatment, and the distinction is worth holding onto.

What This Evidence Does Not Establish

It is laboratory conditioning, not life. A tone paired with a mild shock is a controlled proxy for learning about danger. It is a genuine model and it is not the same as a person’s history.

Pooling studies fixes power, not design. Combining data across studies to reach 601 participants addresses one of the three problems the authors named. It does not make the paradigm more like the world.

The 2026 network study is a comparison, not an intervention. Naming neurofeedback and TMS as possible future targets is not evidence that either changes anything, and should not be read as such.

And none of it measures how anyone feels. ‘Feeling safe’ is a subjective state; what these studies measure is a conditioned response and its neural correlates. The gap between the two is real and this research does not close it.

What it does support is narrow and worth having: safety learning is a distinct process, with partly distinct anatomy and slower timing than threat learning. A nervous system that has been under sustained load is doing something specific when it settles — not merely stopping something else.