The Safety That Doesn’t Travel
State-Dependent Extinction in ADHD

The question

There is an asymmetry at the centre of extinction research that almost nobody outside the field knows about, and it is the most useful fact in this whole area.

Fear memories travel. Extinction memories do not.

Condition an animal to fear a tone, extinguish the fear until it is gone, then test it somewhere else - a different room, different lighting, different smell - and the fear comes back at full strength. This is called renewal, and it is one of the most reliable findings in learning research. ¹ The original fear generalises across contexts. The safety learning stays welded to the place it was acquired. Vervliet and colleagues put it in one line: fear extinction is relatively easy to learn and difficult to remember. ²

Now make the context internal. Contexts do not have to be rooms. A drug state can serve as one, and when it does, extinction learned under the drug fails to retrieve off it. ³

And now put that next to ADHD, which is not a condition of constant impairment but of fluctuation. Under-arousal in ADHD is context-specific rather than fixed, and arousal also becomes measurably more unstable as a session wears on. ⁴ The most reliable cognitive finding in the whole ADHD literature is not a deficit at all - it is variability. ⁵

Put the three together and a question falls out:

Is extinction learning in ADHD state-dependent - acquired in one internal state and failing to retrieve in another?

If it is, then the picture is not somebody who cannot learn that things are fine. It is somebody who has learned it, repeatedly, and cannot reliably get to it. As far as I can find, nobody has asked. ⁶

Where the question came from

It came from a problem with a proposal I had just finished writing.

In Nothing Happened I set out a case that extinction learning is slower in ADHD, built on the finding that dopamine neurons signal the moment an expected aversive event fails to arrive. ⁷ The design was a between-person comparison: are ADHD adults over-represented among slow extinguishers?

Then I noticed what that design cannot see.

In the domain where ADHD has been studied most intensively, a meta-analysis of 319 studies found substantially elevated reaction time variability - and, once variability was accounted for, no slowing of processing speed whatsoever. ⁵ The signature is inconsistency, not decrement. A study asking whether ADHD adults extinguish more slowly on average is asking a between-person question about something that may be a within-person fluctuation, and a between-person design cannot see a fluctuation. It can only average it flat.

That does not make the first question wrong. It bounds what an answer to it would mean. If extinction performance in ADHD moves with state, a single-session between-person study can still detect a deficit large enough to survive the noise - but its null becomes uninterpretable, because “no deficit” and “a deficit averaged flat” produce the same flat line. A study whose positive result counts and whose null settles nothing is worth running while it is cheap, and worth improving on once it is not.

That is not a refinement of the first proposal. It is a different hypothesis with a different mechanism, a different design and a different outcome measure, and it needs its own study.

It also happens to explain something the deficit account cannot. If extinction in ADHD were uniformly impaired, a person’s threat expectancies would be uniformly unextinguished, and the resulting pattern would be legible: everything of a kind stays frightening. That is not what people describe. What they describe is patchy - some rejections release, some never do, and the difference between them does not track how bad they were. A state-dependent account predicts exactly that patchiness, and predicts that the boundary is set by something with no relationship to the content of the event at all.

What the evidence actually says

Extinction is the half that does not travel

Bouton’s account of extinction is the standard one and it is worth stating carefully, because the popular version is wrong.

Extinction does not erase the original association. It creates a second, competing one, and the two coexist. Which of them controls behaviour at any moment is decided by context. Because the fear memory was learned first and without contextual qualification, it retrieves broadly; the extinction memory was learned second, as a correction to an existing expectancy, and it retrieves narrowly - only under conditions resembling those in which the correction was made. ¹

That asymmetry produces three well-documented ways for extinguished fear to come back. Renewal, when the context changes. Spontaneous recovery, when time passes. Reinstatement, when the aversive event is encountered again on its own. All three are routine in humans, not just laboratory curiosities in rodents. ²

This is the fact that makes the present question worth asking. If safety learning is intrinsically context-bound, then anything that reliably changes a person’s context between learning and retrieval will degrade it - and it will do so without any deficit in learning at all.

Internal states work as contexts

The word “context” invites you to picture a room, and that is the wrong picture.

Bouton, Kenney and Rosengard extinguished fear in rats under benzodiazepines and tested them off the drug. ³ The extinction did not transfer. The fear returned. The drug state had functioned as a context in exactly the way a room does - safety learned inside it, unavailable outside it.

That is the load-bearing precedent for this proposal, and I want to be precise about what it does and does not establish. It establishes that an internal, pharmacologically-induced state can serve as an extinction context. It does not establish that an endogenous, self-generated state of arousal or engagement can do the same. That step is the untested one, and it is the centre of this proposal rather than an assumption behind it.

ADHD is a condition of fluctuation, and the fluctuation has been measured

The premise that ADHD arousal is variable is not folk psychology. It has been measured, with the same instrument this study would use.

Du Rietz and colleagues recorded skin conductance in 71 adolescents and young adults with ADHD and 140 controls, across rest, a slow low-demand task, and fast-paced incentive tasks. ⁴ Lower arousal appeared in the ADHD group only during the slow, low-demanding task, and normalised under more stimulating conditions. Arousal also became more unstable towards the end of the session. Their conclusion is the premise of this proposal stated as a finding: under-arousal, and fluctuating arousal, are context-specific rather than stable impairments in ADHD.

That result also hands the study its manipulation. If a slow, monotonous task produces under-arousal in ADHD and a fast, incentivised one does not, then those two conditions are a validated way to move a participant between states.

The theoretical accounts point the same way. State regulation models treat ADHD performance as conditional on task conditions rather than fixed, ⁸ the dual pathway model separates a motivational route from an executive one, ⁹ and reviews of reinforcement effects find ADHD performance disproportionately sensitive to contingency and stimulation. ¹⁰ At the other end of the range sits hyperfocus, which is poorly operationalised but not imaginary. ¹¹

Every one of those describes the same thing from a different angle: the same person performs differently depending on a state that changes on a timescale of minutes to hours.

The one animal result that bears on it

Brackney and colleagues tested extinction in spontaneously hypertensive rats, the most-used rodent model of ADHD, against Wistar-Kyoto controls. ¹² Response bout length shortened progressively during extinction for the controls and did not for the SHR, which the authors read as slower updating of the response-outcome association: an extinction learning deficit.

Two caveats matter and I will not bury either. It was appetitive extinction, not fear. And the SHR model has well-known limitations as a model of ADHD. But it is the closest existing evidence that an ADHD-like phenotype extinguishes abnormally, it points the same way as my first proposal, and I would rather cite it than pretend the animal literature is silent.

Note also that it does not distinguish the two accounts. A deficit and a state-dependent retrieval failure both look like slower extinction if you only measure one session in one condition, which is the general problem this proposal exists to solve.

Rejection sensitivity, and the thing a deficit account cannot explain

The far end of this remains soft, and it was soft in the first proposal too. Rejection sensitivity dysphoria is not in the DSM-5-TR, has no validated measure, and rests on a small qualitative base. ¹³¹⁴ The general construct next to it is well built, ¹⁵ and the established ADHD construct sitting on top of it - emotion dysregulation - is meta-analytically supported and has to be ruled out before anything here counts as a finding. ¹⁶

But there is one feature of the clinical description that a uniform deficit account cannot accommodate and this one can.

The reactions are not uniform. People describe specific rejections that never lost their charge sitting alongside others, of comparable severity, that faded normally. If extinction were simply impaired, that heterogeneity would need an explanation from outside the model - severity, timing, who it was, what it meant. A state-dependent account explains it from inside: what varies is not the event but the condition the person was in when the disconfirming evidence arrived afterwards. Which predicts that the boundary between the rejections that released and the ones that did not should be poorly predicted by anything about the rejections themselves.

That is an odd prediction, it is specific, and I have not seen anyone make it.

The framing

In ADHD, internal state functions as a context for extinction. Safety learned in an engaged state does not fully retrieve in a disengaged one. Extinction is therefore not slow but intermittent and poorly transferred - and a life’s worth of intermittent, poorly transferred safety learning leaves an expectancy that is neither extinguished nor intact.

Three things recommend this over the deficit account in my first proposal.

It predicts the right statistic. A deficit predicts a lower mean. This predicts higher variance and a specific interaction. Given that variability rather than slowing is ADHD’s most robust cognitive signature, ⁵ the second is the better bet on the shape of the effect before any data are collected.

It needs much less borrowed machinery. The first proposal rests on a dopaminergic mechanism it cannot itself measure, and says so. ⁷ Renewal is a behavioural phenomenon that stands on its own evidence. The dopamine account is welcome here if it holds, but nothing in this hypothesis depends on it.

It explains the existing null. Spencer and colleagues ran fear conditioning and extinction in medication-naive ADHD adults and found atypical circuitry with, as reported, unremarkable conditioned responding. ¹⁷ If extinction performance in ADHD depends on state, participants get sorted by whatever state they were in on the day, and a between-group comparison of means washes out. That is not special pleading invented to save a hypothesis; it is a prediction this account makes about any single-session between-groups design, including my own first one.

What this has to survive

Eight problems, and the first two are the ones that could sink it.

The direction of the arousal argument may be backwards. I have been assuming that an engaged state is the good one for extinction. My own earlier reasoning cuts against that. Extinction needs a phasic prediction-error signal to be legible against background dopamine tone, and if engaged states raise that background, the signal becomes less readable, not more. ADHD imaging reports attenuated tonic with enhanced phasic release, ¹⁸ which on a naive reading makes the low-arousal state the better one for signal detection. The escape is an inverted-U - too little tone and there is not enough receptor occupancy to gate plasticity at all, too much and the burst disappears into the background - with ADHD sitting low. But I have to commit to that, say where hyperfocus falls on it, and accept that if the effect runs the other way the hypothesis is not rescued, only reversed. A study that predicts an interaction without specifying its sign is not predicting anything.

Circularity. “Extinction worked because she was engaged; we know she was engaged because extinction worked.” Without state measured independently and before the outcome, this explains everything and forbids nothing. The whole design stands or falls on committing to a state measure in advance and reporting it whether or not the manipulation worked.

The state measure and the outcome measure are the same instrument. This is the awkward one nobody would spot until analysis. Skin conductance is how this literature measures conditioned responding, and skin conductance level is also how Du Rietz measured arousal state. ⁴ If ADHD participants have lower and more fluctuating tonic skin conductance, then a group difference in the dependent variable is partly a group difference in the independent variable. The separation available is that tonic level and phasic response are distinguishable, and the design should use tonic level as the state measure and phasic differential responding as the outcome - but this has to be pre-specified, and a second state measure on a different modality is not optional.

Renewal is normal. Everybody renews. The hypothesis does not predict that ADHD adults show a state-mismatch effect; it predicts a larger one than controls. That is an interaction, interactions need more power than main effects, and this design has three factors. The sample has to be budgeted for the interaction, argued for in advance, and it will be bigger than this literature usually recruits.

More parameters, less falsifiability. The first proposal had one moving part. This has state, phase, and their interaction, and a model with a state-dependent switch can absorb almost any pattern after the fact. The only real defence is preregistration: state the predicted interaction and its direction, and treat its absence as disconfirmation rather than as evidence the manipulation failed.

Endogenous state is not drug state. The precedent for internal context is pharmacological. ³ Whether an ordinary shift in engagement is a strong enough contextual change to gate extinction retrieval is unknown, and it may simply be too weak an effect to detect. That is the most likely way this returns a null.

Medication is a confound and a prediction at once. Stimulants reduce reaction time variability, ⁵ and raise extracellular dopamine. ¹⁹ If the mechanism is state instability, medicated participants should show a smaller mismatch effect. That is a nice prediction and a nasty confound, and it means medication status cannot be a covariate bolted on afterwards - it has to be designed in.

And the construct at the far end is still soft. ¹³¹⁴ The extinction findings would stand on their own; the rejection sensitivity link would not, unless it survives adjustment for emotion dysregulation. ¹⁶ I have also deliberately declined to frame any of this against complex PTSD, which was where the idea first pointed. ²⁰

The design that solves the hard problem

The obvious design manipulates state at extinction between participants, and it is a bad design: it burns the largest source of variance in the study on a between-groups contrast in a population defined by within-person inconsistency.

The better one puts the whole crossing inside each participant, and it works because a person can hold more than one conditioned stimulus at a time.

Condition two stimuli on day one, CS+A and CS+B, both reinforced, plus an unreinforced control stimulus. On day two, extinguish A under one state and B under the other, with state manipulated by task conditions and order counterbalanced. On day three, test both stimuli under both states.

Every participant now supplies all four cells: extinguished-engaged tested engaged, extinguished-engaged tested disengaged, and the two reciprocals. State-dependence appears as a within-person cost of mismatch, measured against that person’s own matched condition. Individual differences in arousal, in skin conductance amplitude, in fear acquisition and in medication all drop out of the contrast that matters, because each participant is their own control.

That is what converts an underpowered three-way interaction into something a project of realistic size can actually test.

What I would actually do

Participants. Adults with ADHD and matched controls. Medication status recorded and, ideally, a within-subject on-and-off arm rather than a covariate. Trauma history and current PTSD excluded, as in Spencer. ¹⁷

State manipulation, borrowed rather than invented. The engaged condition is fast-paced and incentivised; the disengaged condition is slow, monotonous and low-demand. These are the conditions under which Du Rietz and colleagues observed ADHD arousal to differ and to normalise respectively, ⁴ so the manipulation arrives with evidence that it moves the thing it claims to move in this population specifically.

State measurement, independent of the outcome. Tonic skin conductance level as the primary state index, pupillometry as a second modality, and a brief self-report of engagement at fixed points. Reported for every participant in every block, whether or not the manipulation worked.

Conditioning. Standard differential protocol with two reinforced stimuli and one unreinforced, calibrated aversive stimulus, US expectancy ratings alongside skin conductance, following the consensus guidance so the data can join existing datasets. ²¹ Acquisition equivalence across the two CS+ is a precondition; if A and B are not equivalent at the end of day one, the day-three comparison means nothing.

Primary outcome. Differential phasic response at test, as a function of state match versus mismatch, and the group by match interaction.

Table 1
Predicted Pattern Under Three Accounts

	State-dependent retrieval	Uniform extinction deficit	No effect
Extinction within a session	Normal in the engaged state, poorer in the disengaged	Poorer in both	Normal in both
Test under matched state	Comparable to controls	Poorer than controls	Comparable
Test under mismatched state	Disproportionately poorer	Poorer, but no worse than matched	Mismatch cost equal to controls
Group by match interaction	Present, and the finding	Absent	Absent
Within-person variability	Elevated	Not specifically predicted	Not elevated

Note: The three accounts are separated by the interaction rather than by any main effect. Both clinical accounts predict that ADHD participants do worse somewhere; only the state-dependent account predicts that the disadvantage appears specifically when the state at test differs from the state at extinction, and is small or absent when the two agree.

Analysis. Mixed-effects models on trial-level responses, random intercepts by participant, with the group by match term as the test of record. Within-person variability as a secondary outcome in its own right, since it is the measure the deficit account makes no prediction about.

My honest prediction. A mismatch cost in both groups, because renewal is universal. A larger one in the ADHD group, smaller than the story wants. Elevated within-person variability that turns out to be the most robust thing in the dataset. And a real possibility that ordinary engagement shifts are too weak a context manipulation to move extinction retrieval at all, which would be a clean null and worth knowing.

Why this one is on the list

Because of what it would change about how exposure therapy is scheduled, and the change is cheap.

Craske and colleagues already set out an inhibitory learning approach to exposure, and two of its named strategies are conducting exposure in multiple contexts and deliberately introducing variability. ²² The rationale is precisely the one above: extinction memory is context-bound, so vary the context and it retrieves more widely. This is established, and it is already recommended.

What nobody does is treat the patient’s own internal state as one of the contexts to vary.

Exposure happens in a therapy room, at an appointment, with a clinician present and the patient alert, prepared and motivated. That is one internal state, and it is close to the best one the person will be in all week. The situations the learning has to survive are the tired Tuesday, the flat afternoon, the evening after a bad day. If safety learning is state-bound, then exposure conducted exclusively in the alert-and-motivated state is being installed in a context the person is rarely in when they need it.

The prescription that follows costs nothing: run some exposure trials when the patient is flat, bored, tired or under-stimulated, on purpose. That is not a new therapy. It is one more entry on a list of context variations that clinicians are already advised to make.

And it reframes a sentence that clinicians hear constantly and usually treat as a failure of insight. I know it’s fine. I just don’t feel it’s fine when I’m like this. Under a deficit account that is a person who has not learned. Under this one it is a person who has learned, accurately, and cannot reach it from where they are standing. Those are different problems and only one of them is the patient’s fault, which is a distinction worth the cost of a study.

There is a version of this that goes badly, and it is the same one as last time wearing different clothes. “Their safety learning does not stick” is a short step from “they cannot be relied upon to have processed anything”, and that is a sentence that gets used against people in custody hearings and workplace disputes. The same data supports “vary the conditions and the learning holds”, which is an instruction to a clinician. A study in this area should be explicit about which of the two it is for.

Then the smaller thing, which is why I want to know.

Somebody carrying a rejection from years ago, and released from another one no less severe, has usually gone looking for the difference in the events. What made that one worse. What it meant. What it says about them. If this account is right, the difference may not be in the events at all. It may be in what state they happened to be in during the hundred ordinary afternoons afterwards when nothing bad happened and the evidence was quietly available to be taken in.

That is not a comforting answer. But it is a better question than the one people usually torture themselves with, and it is answerable.

⁂

---