The Computational Anatomy of Psychosis

Picture the brain not as a camera or a tape recorder, but as a betting machine. It constantly wagers on what’s out there and what will happen next, then updates its bets when surprise hits. In predictive coding, those updates travel the cortex as "prediction errors," and they’re weighted by precision — the brain’s confidence that a given error deserves attention. Here’s the key, testable link: that precision looks like the gain on the neurons that report error, especially superficial pyramidal cells in the upper layers of cortex. Turn that gain up, and an error shouts. Turn it down, and it whispers. As Borgwardt, Mechelli, and Adams argue, that single dial connects molecular neuromodulation to symptoms in schizophrenia. Hypofunction of N-methyl-D-aspartate receptors and deficits in gamma-aminobutyric acid interneurons in high-level cortex can reduce precision on abstract beliefs, while state-related surges of striatal dopamine can pump precision into low-level signals. The result isn’t random noise — it’s aberrant inference. And it shows up in perception, in eye movements, and in the sense of agency. Let’s make this concrete with birdsong. The team built a hierarchical generative model where a higher-level, slower Lorenz attractor shapes a faster, lower-level attractor that actually sings. Two high-level control variables modulate the lower system’s Rayleigh and Prandtl numbers, which in turn set the chirp’s frequency — in the two to five kilohertz range — and amplitude over time. The model converts the resulting chirps into a synthetic sonogram and runs Bayesian filtering to infer the hidden causes and states that best explain the sound. When the balance of precisions is set to a healthy baseline — sensory log precision at 2, first-level hidden states at 8, and second-level prediction errors at 16 — the system homes in on the true hidden states after a few hundred milliseconds and predicts the sound accurately. The biggest error signals pop at chirp onsets, with the third chirp provoking a longer error burst, because it’s trickier to forecast. That pattern is exactly what many event-related potential studies interpret as the brain’s response to surprise: a spike when the world deviates from the current story. Now remove the last three chirps — an omission paradigm. Two things jump out. First, at the moment the missing chirp should have happened, you still get a strong error at the sensory level. There’s no input, so the error is purely top-down: a prediction slamming into silence. Second, a brief illusion appears. The model "hears" a faint chirp at the right time but at too low a frequency. Timing is preserved; content is off. That’s a tidy computational echo of omission responses seen in human electroencephalography: the brain’s predictive machinery can generate a phantom of the expected event, tethered to time but not fully formed. Here’s where precision earns center billing. Lower the high-level precision — drop that second-level log precision from 16 to 2 — and two consequences follow. The omission response shrinks; top-down predictions lose their grip, and bottom-up input gets relatively louder. And that third chirp, the one that relies on stable beliefs about the song’s structure, becomes effectively unpredictable. The model fails to perceive it as a coherent event, and the resulting error shows up in a different place than before. That shift mirrors reduced mismatch negativity, the brain’s "something’s off" signal, reported in trait-like features of psychosis. One more twist makes the picture sharper. If, in an attempted compensation, you also reduce sensory precision — drop it from 2 to minus 2 — the omission response vanishes, and tracking collapses. Frequency and syntax unravel even though the tempo roughly persists. The brain is now hallucinating: top-down predictions dominate because the bottom-up anchor has been cut. Errors persist, but because the system treats them as imprecise, their effective amplitude is down-weighted. The story that emerges is simple and powerful. Weaken priors, and everything becomes a surprise; weaken the senses, and predictions fill in the world. Shift from sound to sight, and you find the same logic running eye movements. Adams, Perrinet, and Friston studied smooth pursuit — that lovely, nearly reflexive glide your eyes make when they follow a moving target — through the lens of active inference. In their model, both gaze and target are pulled toward a hidden attractor in visual space, an internal set-point that encodes "we’re looking where the thing is." The retina contributes exteroceptive input about where the target sits, via seventeen channels with Gaussian receptive fields. Proprioception contributes where the eye actually is. And an occluder function — think of it as a light switch that flicks off when the target passes behind a bar — zeros out retinal input during occlusion. Hidden states at the first level track eye angle and velocity; a second level carries a belief about the target’s periodic motion, including its frequency. The first level’s log precision is set around three, while the higher level starts lower, at about minus one. Now we fiddle the same dial: we reduce precision at the higher level. During occlusion, pursuit degrades. When the target re-emerges, the eye lags further behind than it would under normal precision, because the system put less trust in its own high-level forecast while the senses were dark. Once the target is in view, though, tracking is fine, with stable pursuit during stretches around 1.2 to 1.4 seconds and again between 2.0 and 2.2 seconds. If the eye falls too far behind, catch-up saccades kick in, with velocities clearing about 30 degrees per second. So far, so intuitive: down-weight your long-range prediction, and you suffer when the lights go out. But then comes a paradox. Remove the occluder. Shorten the target’s period to about half a second, and have it suddenly reverse course around 780 milliseconds — a true surprise. In that scenario, the model with reduced high-level precision actually pursues slightly better. The lag between eye and target is smaller, and the peak velocity of the compensatory movement is a touch lower. It’s a small effect, exactly as Hong and colleagues have reported in humans, but it makes a deep point. Strong priors stabilize you when the world is regular; the same priors can slow you when the world flips. Precision is a double-edged sword. Agency feels like a different question, but under the hood, it runs on the same machinery. In a compact somatosensory model that Adams and Friston showcase, a single hidden state represents self-generated force. That internal force drives both proprioceptive channels — your sense of where your limbs are — and somatosensory channels — the pressure or touch on the skin. External force drives only the somatosensory side. That asymmetry is crucial because it makes somatosensory evidence ambiguous: the same sensation could be me pushing or the world pushing on me. Active inference solves that by attenuating the precision of sensory signals during self-movement. A parameter, call it g, governs how strongly sensory precision is turned down as your internally generated force rises. Lower precision on prediction errors means your proprioceptive predictions can be fulfilled through reflex-like pathways without being constantly vetoed by your own sensory chatter. With normal attenuation, the model reproduces the classic force-matching illusion. When people are asked to replicate an externally applied force with their own finger, they consistently overshoot. In simulation, internally generated force comes out higher than the external force at every level tested, just as Shergill and colleagues reported. The model even uses a 90 percent confidence interval as a stand-in for the perceptual bound within which a match is considered acceptable, and the overshoot sits comfortably inside that range. Now turn attenuation down. Decrease g from 6 to 2, and the balance flips. Sensory prediction errors get too much weight, proprioceptive predictions are continually contradicted, and movement stalls — akinesia in the model’s terms, a catatonia-like state. There’s an obvious fix: raise high-level precision so that top-down beliefs about internal causes dominate again. In their simulations, boosting the log precision on hidden states and causes by four units restores movement. But it does so at a price. The force-matching illusion disappears, and a new misattribution enters: the system infers an antagonistic external cause that mirrors the internal one. In plain language, "something else is pushing back," even when the source is you. That’s a mechanistic route to a somatic delusion. Two empirical anchors make this more than a thought experiment. Teufel and colleagues found that healthy people with high delusional ideation show a diminished force-matching illusion, and Shergill’s patient studies show the same attenuation in those with prominent positive symptoms. Less sensory attenuation, less illusion; compensate with strong priors, and agency starts to slip. Across these vignettes — birdsong omissions, occluded pursuit, and the feel of self-made force — the same geometry of inference appears. Precision-weighted errors rise from the senses through superficial pyramidal cells; predictions descend; and the balance between them shapes perception, action, and belief. When high-level precision is too low, the world looks surprising and fragile. When sensory precision is too high, your own movements feel intrusively loud. When dopamine in the striatum elevates the precision of certain prediction errors, as many antipsychotic-sensitive states seem to reflect, the system can be driven into hypervigilant, state-like modes. Meanwhile, trait-like features — linked by Borgwardt, Mechelli, and Adams to cortical N-methyl-D-aspartate receptor hypofunction and GABAergic deficits in supragranular layers of prefrontal and medial temporal areas, with possible involvement of D1 receptor signaling — can keep the gain turned down at the top of the hierarchy. That dissociation helps explain why antipsychotic drugs, which mainly target D2 receptors, shift states but don’t necessarily normalize traits. If you want one line that unifies this, it’s this: psychosis is not a single lesion; it’s a disorder of precision. Superficial pyramidal-cell gain implements that precision in the cortex. Dynamic causal modeling studies in schizophrenia already point the same way, showing patterns consistent with decreased high-level precision and increased low-level precision in sensory circuits. Put differently, the bets are being placed at the wrong table, with the wrong stakes. There’s something satisfying about how these models line up with lived phenomena. Reduced mismatch negativity in at-risk states maps onto weakened high-level priors in the birdsong. Slightly better pursuit on true surprises, under lower precision, explains those counterintuitive human data. The presence or absence of the force-matching illusion becomes a readout of how much you attenuate your own sensory feedback during action. And the sudden, eerie feeling that your body is being moved by something else slots into the same framework when the compensation for weak attenuation overshoots. Where does that leave us? With a playbook for testing and, maybe, treating. Because the claims are computational, they’re falsifiable. You can perturb precision pharmacologically — nudging N-methyl-D-aspartate or dopamine — and watch how omission responses, pursuit under occlusion, or force matching shift. You can use model-based analyses of brain signals to estimate precision at different levels and track how those estimates change with symptoms or drugs. And you can design behavioral tasks, like targeted omission paradigms or pursuit with brief unpredictable reversals, that read out the balance between priors and sensory evidence. The broader promise is restraint and specificity. Rather than guessing at circuits piecemeal, you ask: where is precision misallocated, and can we turn that dial? That’s a modest goal. But as these studies from Borgwardt, Mechelli, Adams, Perrinet, and Friston show, it’s a goal that ties together molecules, layers, rhythms, and misbeliefs with a single, intelligible thread: how confident the brain is in its own errors, and when it should be.