Periaqueductal gray (PAG)
The structure the wiki has been pointing at from four directions without ever putting it in one place.
What it does, from the wiki’s existing material
- The output stage of defence. survival-circuits and LeDoux (2012) place it as the effector for freezing, flight and defensive autonomic mobilization — a coordinated behavioural package, not a feeling. tonic-immobility and stress-induced-analgesia are its columnar specialities, and van der Kolk’s opioid-mediated analgesia under trauma reactivation runs through it.
- The bottom of the visceromotor hierarchy. Petzschner et al. (2021) specify that interoceptive active-inference “requires explicit neuronal representation of predictions and errors (visceromotor hierarchy down to hypothalamus/PAG/parabrachial).” computational-psychiatry inherits the same architecture: prediction errors determine regulatory signals sent from visceromotor regions and brainstem structures like the PAG out to the autonomic nervous system (Stephan et al. 2016).
- A respiratory anticipation/experience dissociation. respiratory-interoception records the wiki’s one anatomically-addressed anticipation/experience split: ventrolateral versus lateral PAG for the anticipation versus the perception of respiratory threat (Faull et al. 2015, 2016; Faull & Pattinson 2017). Those papers are Harrison’s own earlier work under her former name.
- Pain and placebo. PAG activity has been related to aversive prediction errors in pain (Roy et al. 2014) and to the precision of prior beliefs in placebo analgesia (Grahl, Onat & Büchel 2018) — the two results that made a computational reading of this structure plausible in the first place.
What is new: it carries error, not certainty
Harrison et al. (2021) is the wiki’s first source to regress human PAG activity on a fitted trial-by-trial interoceptive quantity, and the dissociation is clean:
| quantity | PAG |
|---|---|
| prediction error magnitude (breathing) | activation |
| prediction certainty (breathing) | nothing |
Compare the cortical picture in the same slab, where anterior insula, ACC, MFG and dlPFC all tracked both — deactivating with certainty and (aIns, ACC, MFG) activating with error.
So the cortex holds both sides of the update equation and the midbrain holds one. That is exactly the division of labour the hierarchical account specifies: beliefs and their precision live in cortex; the brainstem receives the deviation and acts on it. It is the wiki’s first direct evidence for the lower half of an architecture it has otherwise held on cytoarchitecture and theory.
Two cautions on how far to push it. The PAG did not show the valence effect the pain literature might predict, and it showed no anxiety-group effect at all. And a null for certainty in a small midbrain structure imaged even at 7T is a weak null — the authors’ own reason for using a reduced field of view centred on insula and midbrain was that this structure is hard to resolve.
Why it complicates “where feelings are constituted”
The PAG is the sharpest test case for where-are-feelings-constituted. It produces the full defensive package — freezing, analgesia, autonomic mobilization, altered breathing — without cortex, which is Panksepp’s and LeDoux’s shared datum read to opposite conclusions: evidence that feeling is subcortical, or evidence that a coordinated survival response is not a feeling at all.
The Harrison result does not settle that and slightly reframes it. If the PAG’s activity here is best described as prediction error — a deviation signal defined relative to a model held elsewhere — then what it carries is not a state to be felt but a discrepancy to be corrected, and the quantity that would need to be felt (the belief, and its certainty) is cortical. That reading favours LeDoux’s side of the argument, and it is an inference from a regressor, not a demonstration.
Open
- The columnar organization (dorsolateral/lateral/ventrolateral, each with its own defensive strategy) is central to the animal literature and invisible in every human result the wiki holds, including this one.
- The anticipation/experience dissociation from Faull et al. and the error/certainty dissociation from Harrison et al. are two different cuts through the same structure by overlapping authors, and nobody has run them together.
- No wiki source measures PAG activity alongside any interoceptive accuracy measure, so its relation to perception rather than to regulation is untouched.