a/research· 31 October 2026 · 7 min read

Whose belief is your diagnosis? A dead fish, 442 patients and one safe choice

Cover of The Lonely Whisper of the Wind (Omxus Press)

Eleanor Longden was a university student when a voice began narrating her day. Within two years she had the whole frenzied repertoire, as she later put it, and a diagnosis of schizophrenia. Her psychiatrist told her: "Eleanor, you'd be better off with cancer, because cancer is easier to cure than schizophrenia." In her words, "I'd been diagnosed, drugged and discarded" (Longden, 2013). To see what went wrong, it helps to meet a fish.

The salmon in the scanner

In 2009 a team of researchers put an Atlantic salmon in an MRI scanner. Their write-up is exact about the subject: about 18 inches long, 3.8 pounds, and "not alive at the time of scanning". They showed it photographs of people in social situations and asked it what each person was feeling. Then they analysed the scan the way many published studies did at the time.

Clusters in the dead fish's brain lit up as statistically significant (Bennett et al., 2009).

The fish wasn't thinking. The statistics were. Test enough tiny boxes of brain at a loose threshold and some will pass by chance. Seven years later a bigger test showed the joke had teeth: on scans of healthy people at rest, standard methods found "activity" up to 70% of the time when they promised 5% (Eklund et al., 2016).

A brain that predicts

Hold the fish in mind, because psychiatry now uses a similar engine on people. A popular theory says the brain is a prediction machine. It guesses what comes next, compares that with the senses, and updates (Friston, 2010). The guess before the evidence arrives is called a prior.

In this language, autism has been described as priors that are too weak, a world that feels "too real" (Pellicano & Burr, 2012), and psychosis as predictions weighted wrongly (Sterzer et al., 2018). These are serious ideas from serious scientists. Each one, though, needs a standard: the prior a healthy brain should hold.

Who knows what your brain should expect, given a life nobody else lived?

Same patients, different answers

A conclusion of this kind is a prior multiplied by data. When the data are thin, the prior decides. You can see how much. Take a public dataset of 442 diabetes patients (Efron et al., 2004). In the full group, a higher body-mass index clearly goes with faster disease progression. Now look at the same patients 25 at a time, 200 times over.

  • A prior with no strong opinion finds "no clear effect" in 24.5% of samples, which is simply the power of 25 patients.
  • A sceptical prior centred on zero finds "no effect" in 96.0% of them.
  • A confident wrong prior says body-mass index slows the disease, in every single sample.

Nothing about the patients changed between rows. Only the belief brought to them changed. Even with all 442 patients, the sceptical prior still shrinks the effect by 41%.

One person, one afternoon

A mental-health assessment is the smallest sample there is: one person, one afternoon. If 25 patients can't outvote a confident prior about something as measurable as weight and disease, one person can't outvote a clinician's frame about what their eye contact means.

A small simulation shows the shape of it. Give three simulated classifiers the same 500 behaviour profiles, and 54.2% of the people get opposite verdicts: "pathological" under one frame, "adaptive" under another. Whoever built that simulation set how far the frames disagree, so the number isn't the lesson. The lesson is that a classifier can only return the categories it was given, and "what happened to you?" isn't a category in the pathology frame.

That matters, because the behaviours on the checklist (less eye contact, flat affect, slow answers, a story told out of order, sensitivity to noise and touch) are also what people do after threat. Childhood adversity raises the odds of psychosis almost threefold (Varese et al., 2012).

Longden found this out the long way. Her psychiatrist eventually gave her a leaflet for the Hearing Voices Network, and with those people she came to see her voices as "a meaningful response to traumatic life events, particularly childhood events". She came off her medication gradually, and ten years after the first voice she graduated with the highest degree in psychology her university had given. One voice, she says, dictated the answers in an exam, "which technically possibly counts as cheating". Her argument is one line: "an important question in psychiatry shouldn't be what's wrong with you but rather what's happened to you" (Longden, 2013). That's one person's story, and medication saves plenty of others. But nothing about her changed, except the prior.

Nobody can compute it anyway

Even with the right prior, the sums are out of reach. Exact reasoning of this kind is provably hard: the work explodes as the network grows (Cooper, 1990), and cutting corners doesn't escape the proof (Kwisthout et al., 2011). A brain has about 86 billion neurons (Azevedo et al., 2009). Nobody, brain or psychiatrist, is computing the exact answer. When a model calls a person's inference "aberrant", it compares that person with an idealised computation no one performs, under a prior no one measured.

The choice that's right every time

So what should anyone do when the prior can't be known? Statisticians have an answer: choose the action that works under every prior you can't rule out (Berger, 1994).

For a person in distress, that action is safety. Food, sleep, no threats, someone calm, someone who believes you. It doesn't need to know your history, so it can't be wrong about it. If you've got a brain disease, safety helps. If you've got a terrible past, safety helps. A diagnosis-first approach needs its prior to be right. Safety-first doesn't. (A toy model of 200 agents with hidden histories shows this, though its authors wrote its rules, so it illustrates the logic and proves nothing about clinics. Parts of the polyvagal theory often quoted here are also disputed, and this argument doesn't need them; Grossman, 2023.)

The clinicians who use these models aren't villains. They believe they're helping, and most of them do. The trouble is a method that never tells its user how much of the answer they brought with them.

What you can do

Almost everyone gets it wrong about someone they love at some point, and we can start by holding what we think a little more loosely.

  1. If someone you love is in distress, make them safe first and work out the name for it later. Ask what happened. Then listen.
  2. If you've got a diagnosis that doesn't feel right, you're allowed a second opinion. Ask the clinician: "What would make you change your mind?" Please don't stop medication alone or suddenly; plan any change with a doctor.
  3. If it's urgent, Lifeline is on 13 11 14, any hour. In an emergency, call 000.

A dead fish lit up because the method expected something to. You aren't a dead fish. You're allowed to ask who was doing the expecting.

Sources

Azevedo, F. A. C., Carvalho, L. R. B., Grinberg, L. T., Farfel, J. M., Ferretti, R. E. L., Leite, R. E. P., Jacob Filho, W., Lent, R., & Herculano-Houzel, S. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology, 513(5), 532–541. https://doi.org/10.1002/cne.21974
Bennett, C. M., Miller, M. B., & Wolford, G. L. (2009). Neural correlates of interspecies perspective taking in the post-mortem Atlantic salmon: An argument for multiple comparisons correction. NeuroImage, 47, S125. https://doi.org/10.1016/S1053-8119(09)71202-9
Berger, J. O. (1994). An overview of robust Bayesian analysis. Test, 3(1), 5–124. https://doi.org/10.1007/BF02562676
Cooper, G. F. (1990). The computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence, 42(2–3), 393–405. https://doi.org/10.1016/0004-3702(90)90060-D
Efron, B., Hastie, T., Johnstone, I., & Tibshirani, R. (2004). Least angle regression. The Annals of Statistics, 32(2), 407–499. https://doi.org/10.1214/009053604000000067
Eklund, A., Nichols, T. E., & Knutsson, H. (2016). Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates. Proceedings of the National Academy of Sciences, 113(28), 7900–7905. https://doi.org/10.1073/pnas.1602413113
Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
Grossman, P. (2023). Fundamental challenges and likely refutations of the five basic premises of the polyvagal theory. Biological Psychology, 180, 108589. https://doi.org/10.1016/j.biopsycho.2023.108589
Kwisthout, J., Wareham, T., & van Rooij, I. (2011). Bayesian intractability is not an ailment that approximation can cure. Cognitive Science, 35(5), 779–784. https://doi.org/10.1111/j.1551-6709.2011.01182.x
Longden, E. (2013). The voices in my head [Video]. TED. https://www.ted.com/talks/eleanor_longden_the_voices_in_my_head
Pellicano, E., & Burr, D. (2012). When the world becomes "too real": A Bayesian explanation of autistic perception. Trends in Cognitive Sciences, 16(10), 504–510. https://doi.org/10.1016/j.tics.2012.08.009
Sterzer, P., Adams, R. A., Fletcher, P., Frith, C., Lawrie, S. M., Muckli, L., Petrovic, P., Uhlhaas, P., Voss, M., & Corlett, P. R. (2018). The predictive coding account of psychosis. Biological Psychiatry, 84(9), 634–643. https://doi.org/10.1016/j.biopsych.2018.05.015
Varese, F., Smeets, F., Drukker, M., Lieverse, R., Lataster, T., Viechtbauer, W., Read, J., van Os, J., & Bentall, R. P. (2012). Childhood adversities increase the risk of psychosis: A meta-analysis of patient-control, prospective- and cross-sectional cohort studies. Schizophrenia Bulletin, 38(4), 661–671. https://doi.org/10.1093/schbul/sbs050

Keep reading