A hundred friendly dogs can leave a phobic person more afraid than before.
The common-sense cure for a fear of dogs is to meet a lot of friendly dogs. Each calm encounter should be a small piece of evidence that dogs are safe, and the evidence should add up. Sometimes it does. But sometimes the person who is flooded with friendly dogs comes out more terrified than before. How can evidence against a belief make the belief stronger?
Scott Alexander’s answer is that the belief has become a trapped prior: strong enough to consume its own evidence. The mechanism is not willful denial. It is that the prior reaches into the raw experience itself and rewrites what arrived before the conscious mind gets a chance to evaluate anything.
How the prior eats the dog
In the predictive processing account, every perception is a negotiation between a top-down prior (what the brain expects) and a bottom-up sensory channel (what the senses report). A weighting algorithm combines them, and the output is what you experience as reality. Most of the time the negotiation is fair, and you experience something close to what is actually there. Beliefs update when bottom-up surprise is large enough to force the prior to revise.
A trapped prior is what happens when the weighting assigns the bottom-up channel a coefficient near zero. The prior wins every match. So the phobic person does not see a friendly dog and then explain it away. By the time the dog reaches conscious awareness it has already been repainted in the colors of every previous fear, and what they experience is the prior re-rendered with the new stimulus as a prop. That experience, already shaped by the prior, is then fed back as the next datum. The loop runs: prior, distorted perception, stronger prior. Each friendly dog becomes one more terrifying dog, and there is no point at which the truth could leak in, because every input passes through the filter it would need to update.
The same architecture explains why more information makes a trapped person wronger, not righter. Studies that find more scientifically literate people hold more polarized views are not paradoxes on this model. A partisan handed a clean fact about the other side reads it as fresh proof of malice, because the fact is processed top-down and a stronger prior turns each new fact into a sharper instance of itself. At the extreme, the disliked person can do nothing without it counting against them: eating crackers becomes evidence, and so does silence.
Why fear closes the channel
What pushes the weighting toward zero? It seems to track emotional intensity: the brain narrows bottom-up bandwidth in proportion to threat. That is presumably adaptive, since a prey animal that reprocesses every rustle as raw data dies the first time the rustle is real, and trusting the prior is cheaper than trusting a noisy world. But the shortcut creates the trap. Trauma victims often describe a violent event in oddly short, undetailed sentences, because perceptual bandwidth was throttled at the time and the memory has the resolution it had on intake. The window of tolerance names the clinical version: outside the window, the prefrontal mediation that would let bottom-up signal compete with the prior goes offline.
So the dog makes the phobic person afraid, the fear narrows the channel, and the narrowed channel guarantees the next dog will look frightening too. This is why you cannot reason a phobic person out of a phobia, an abuse survivor out of a flinch, or a tribal partisan out of contempt by handing them facts. The facts arrive through a channel that has been narrowed to almost nothing.
Not every confident belief is trapped, though. A strong prior has a high weight but will still update under sufficient surprise; a trapped prior’s weight on the bottom-up channel has collapsed, so no amount of surprise reaches the update step. The difference shows at the boundary, and it suggests a test: if you cannot describe the observation that would change your mind in concrete detail, the prior is probably trapped, not just confident. Confidence can name its falsification condition. Trappedness has a rationalization ready for every imaginable input.
Widening the channel
If the trap is in the perceptual channel, the escape has to be there too. Argument, evidence and exhortation all assume an open channel, so they cannot help. What helps forces a window in which the prior loses its grip on perception.
For the dog phobia, that means graduated exposure with regulation. Exposure alone, with the prior intact, produces another distorted experience that strengthens it. The protocol that works pairs exposure with bottom-up regulation (breath, grounding, a calm therapist’s nervous system co-regulating yours) so that some signal reaches awareness intact. The dog has to be experienced as a dog, even briefly. One clean perception leaks the loop.
For beliefs, a similar move is mechanistic reasoning. Asking someone to explain how their belief actually produces the consequences they expect, rather than to assert it, is one of the few interventions shown to reduce political extremism. Articulating a chain of cause and effect requires holding the prior aside and looking at the world’s actual joints. Psychedelics may work on the same principle chemically: classical serotonergic drugs appear to lower the precision weighting on top-down models, opening a window in which years of un-integrated counter-evidence can land at once.
Together these cases show that trapped priors are not a failure of rationality applied to perception — they are a malfunction of perception itself. Telling a phobic person to “just look at the dog” asks them to use a tool the prior controls. An open channel is a prerequisite for any belief revision, and trapped priors are exactly the configuration in which the channel is not open. It also changes what disagreement asks of you. If someone’s position sits inside a trapped prior, arguing louder raises the threat signal and narrows the channel further, so most fights trap both priors more firmly than before.
Alexander noted, only half-jokingly, that his own prior on the theory had become trapped: every new domain he applied it to confirmed it. The model predicts that too. Which leaves a question worth asking of any idea that explains everything you look at: is it seeing the dog, or painting it?
References:
- Scott Alexander, Trapped Priors As A Basic Problem Of Rationality, Astral Codex Ten
- Andy Clark, Surfing Uncertainty
- Sloman & Fernbach, The Knowledge Illusion — mechanistic reasoning as a depolarizing intervention