Two people can share a diagnosis without sharing a single symptom.
"Mental illness is a brain problem" — almost everyone says this now. Push one step further and it gets awkward. What does a depressed brain look like on a scan? There is no such scan. Is there a blood test for anxiety? No. After decades and billions, psychiatry still has not one biological measure you can use to make a diagnosis. The usual explanation is that the technology isn't good enough yet. The more convincing explanation these days is different: we have been hunting for lesions with a map whose grid lines are in the wrong place. It isn't that there's no mechanism in the brain — it's that the disease names themselves may not sit on any natural seam.
How does the diagnostic manual define major depression? List nine symptoms; if five of them are present for two weeks and at least one is a core symptom, that's a diagnosis. Sounds rigorous — until you count the combinations. A few hundred different symptom sets satisfy that rule. So this is entirely possible: one person can't sleep, can't eat, can't sit still, blames themselves, wants to die; another sleeps too much, eats too much, moves slowly, is exhausted, can't hold attention. Not one symptom in common, and the same diagnosis on both charts.
Comorbidity is the rule, not the exception: of people who meet criteria for one psychiatric diagnosis, close to half meet criteria for a second and a third. If these were genuinely separate diseases, that number makes no sense. It looks much more like one underlying thing surfacing from different sides. Some researchers took every symptom from thousands of people and factored them statistically; the largest shared component ran across nearly every diagnosis. They call it the p factor — roughly, "how dysregulated this person is overall".
So "we can't find the lesion of depression" may not be a microscope problem. Pack a group of mechanistically different people into one box, then search that box for a common brain signature, and you should expect to find nothing — averaging is exactly what erases it.
There's a familiar counterpart in AI: benchmarks only see outputs. Two models score the same on one test — one actually reasons, the other memorised the answers during training. Classify by behaviour alone and you can never separate those two. The whole point of the recent interpretability work is to open the box and re-sort systems by internal mechanism. Psychiatry is stuck on the same problem, except its box is a living brain and you don't get to open it.
In 2013 the then-director of the US National Institute of Mental Health did something startling: he announced the institute would stop organising its research funding around the manual's categories. The reason is the one above — group by disease name and you collect a mechanistically mixed crowd.
The replacement is called RDoC, and you can think of it as a change of coordinates. It doesn't start with "what disease is this?" It asks where a person sits along a handful of functional dimensions: response to threat, response to reward, cognitive control, social processing, arousal and rhythm. Each dimension is then studied across several levels at once — genes, molecules, cells, circuits, behaviour, and what the person reports. So the study group stops being "depressed vs controls" and becomes "everyone whose threat response is over-reactive, whatever the manual says they have".
A decade on, the ledger deserves an honest reading. That same director later said the institute published a great many excellent papers during his tenure, and did not move the needle on suicide, hospitalisation or recovery. RDoC has not become a diagnostic tool anyone uses in clinic, and critics say it promoted "circuits" to an answer before it had earned the promotion. A more conservative alternative, HiTOP, assumes no neural mechanism at all: it just lets statistics sort symptoms into a hierarchy running from specific to general. The two camps argue plenty, but they agree on the sentence that matters — the grid needs redrawing.
"Circuit dysregulation" sounds abstract. Concretely it means three things: gain, balance, switching.
Gain is how excitable a patch of neurons is overall. Everything the brain does is held in place by a pair pulling against each other: glutamate presses the accelerator, GABA presses the brake — the two main chemical messengers (neurotransmitter systems). On the brake side, one class of small inhibitory neurons matters especially: in people with schizophrenia, markers of these cells are repeatedly found reduced, and the fast rhythms in their EEG are weaker. Weaken the brake and the signal-to-noise ratio of the whole region collapses — not because the signal shrinks, but because the background noise is no longer held down.
Balance and switching are about whole networks. You've met two of the three: the default mode network (most active when you're daydreaming or thinking about yourself) and the central executive network (on duty when you're doing something). The third is the salience network, and its job is that of a switch operator — it decides what deserves to be taken seriously, and which network you should be on right now.
The rumination of depression is a failure to switch out: the moment to leave the inward network arrives and you can't leave. And the core abnormality of schizophrenia is increasingly pinned on the switch operator tagging the wrong things: dopamine's problem isn't that there's too much of it, it's that at the wrong moments it stamps "important" onto things that aren't. A parked car, a phrase on television — everything seems loaded with meaning. Delusions are often the explanation built afterwards: a person working hard to find a story that makes those forcibly highlighted experiences hang together.
One layer down there's an account that sits even closer to Topic 1. A healthy brain, at the moment it generates a thought of its own, also issues a copy tagged "this one came from me" — that's how internal products stay separated from external input. Let that tag fail, add top-down expectations that are over-trusted and sensory evidence that is over-ruled, and your own inner monologue gets labelled as coming from outside. That's one mainstream mechanistic account of hearing voices. Worth noting alongside it: healthy volunteers given a drug that blocks glutamate receptors (the ketamine family) briefly reproduce a strikingly similar set of experiences — hard evidence that this was never only about dopamine.
Two tight ones. First: a model invents a passage, then feeds its own invention back into its context as though it were something retrieved — unable to distinguish "I generated this" from "this came from outside", which is precisely the missing self-tag above. Second: when a model's output falls apart, it's rarely the weights that broke; usually it's the knobs governing how much to trust its own priors versus the evidence in front of it (guidance in image models, temperature in chat models). That is exactly what circuit dysregulation says: every part is present, the parameters are off — and it's why psychiatric treatment looks more like tuning than like replacing a part.
Once illness looks like mis-set parameters rather than a broken part, the treatment logic shifts with it.
Step one: the target is computed, not picked off an anatomy chart. Transcranial magnetic stimulation (a magnetic field driving a small patch of cortex from outside the skull) was originally aimed at a fixed spot on the prefrontal cortex. Then came the finding that the more anti-correlated that spot is with a small deep region tied to depression (the subgenual cingulate), the better the response — so the target can be set from each person's own connectivity map. Combine that idea with a much denser stimulation schedule and small trials have reported very high remission rates, along with regulatory clearance. The samples are still small; don't file it under solved.
Step two: closing the loop. Deep brain stimulation used to mean an implanted electrode delivering current continuously, with the dose adjusted by a human every so often. A single case in 2021 did something else: first find, in this particular brain, a pattern of electrical activity that recurs whenever mood drops; then use it as a trigger, and deliver current only when it is detected. That one change turns treatment from "dosing on a schedule" into control with feedback.
Step three is in the pharmacy. An antipsychotic approved in 2024 is the first in decades that doesn't work by blocking dopamine receptors: it acts through acetylcholine instead, pulling dopamine signalling back into range indirectly, with a noticeably different side-effect profile as a result. It's worth singling out because it is itself a product of circuit thinking — not topping up some "level", but entering from elsewhere to move the operating point of a whole chain.
The limits belong in the same paragraph. All of the above is still expensive, small-scale, and available to few. The famous study that used brain imaging to split depression into subtypes failed independent replication; the consensus now is that brain-behaviour associations need samples in the thousands before they stabilise. So the honest position is this: the mechanistic map is being redrawn, and there is still no clinical ruler that tells you which subtype you are and which treatment to use. If you or someone near you is stuck in this, don't wait for that ruler — the average effect of existing treatment isn't dazzling, but it beats doing nothing, and getting a professional involved is the one step that genuinely speeds things up.
"The entity isn't there when you look for it" is a problem other traditions have chewed on for a long time: