The part that makes you see things and the part that makes you better may not be the same part at all.
For decades the argument about psychedelics has been stuck between two camps: a door onto some deeper truth, or just a chemical scrambling the brain for a few hours. The last decade of neuroscience offers a third and much colder version — a receptor called 5-HT2A gets turned, the cortex's pecking order is temporarily flattened, and then a plasticity window stays open for weeks. Which sets up the sharpest question in the field: are the visions the reason it works, or just the noise along the way?
LSD, psilocybin (the active ingredient in "magic mushrooms"), DMT, mescaline (from certain cacti) — these molecules look nothing alike, and their effective doses differ by hundreds of times. Yet the experiences people report are strikingly similar. In pharmacology that's a strong hint: they are probably acting in the same place.
That place is the 5-HT2A receptor. A receptor is just a keyhole on a cell — a molecule of the right shape slots in, and only then does the cell respond. This particular keyhole normally belongs to serotonin, a chemical messenger that mostly sets the brain's background tone rather than issuing direct orders.
How do we know the lock is the switch, rather than something that merely happens to be nearby? There's a clean experiment. Give someone ketanserin — a drug that blocks 5-HT2A and largely leaves everything else alone — and then give them LSD. The subjective effects are almost entirely wiped out. Plug the keyhole and the key is worthless. That's causation, not correlation.
One more coincidence of location is worth noticing. Where 5-HT2A is densest in the cortex is on layer-5 pyramidal neurons — the cells that wire regions together and broadcast predictions back down toward the senses (the thread from Topic 1). The drug isn't stirring the pot at random. It turns one very specific screw, at the top of the hierarchy. (The messenger system itself: Serotonin system)
Which raises an obvious question: your brain is full of serotonin every day, and serotonin opens this same lock. So why aren't you tripping? Hold that thought until section three.
The most widely repeated claim — that psychedelics let you use parts of the brain you normally don't — is backwards. What imaging actually shows is the opposite: something gets turned down.
The first thing noticed was the default mode network: the set of regions most active when you're drifting, thinking about yourself, narrating your own life. After a dose, its internal synchrony drops. A carefully designed 2024 study made the picture much sharper: the same people received 25 mg psilocybin and, on another occasion, methylphenidate (a common stimulant, used as an active control). Psilocybin disrupted functional connectivity more than three times as much — and the mechanism wasn't "this area up, that area down." It was desynchronisation: the correlations within networks and the anticorrelations between them were erased together, so the boundaries between networks simply dissolved. The default mode network was hit hardest, and its suppressed link to the anterior hippocampus was still not fully back weeks later.
Why would falling walls produce that kind of experience? Pick up the thread from Topic 1 and it follows: what you see = the brain's prior expectation × the evidence in front of you, weighted by which one is more reliable. What psychedelics do, in Carhart-Harris's framing, is REBUS — relax the weight on high-level priors. Loosen the priors and everything they normally hold down comes up: the wallpaper starts breathing (low-level edge signals no longer corrected away by "walls are flat"), a passing remark feels world-historically significant (runaway meaning). And when the topmost prior of all — "I am a continuous, bounded person" — loosens too, you get what's called ego dissolution (Topic 17 approached the same thing from the other side). (The network itself: Default mode network)
That step also hands us a plausible reason it might help depression. Depression looks a lot like a state of priors locked shut — the model of yourself (I'm useless, I won't get better) has stopped accepting correction from evidence. Relaxing priors is oil on that lock.
This is the same thing as a knob in generative AI. Image models have a parameter called guidance; chat models have temperature. Both control the same trade: how much to lean on learned regularities versus how much surprise to allow. Turn it toward less constraint and output becomes novel, strange, wildly associative — turn it further and the model confabulates with total confidence. Novelty and nonsense share one knob; there is no setting that buys the first without the second. That is probably also why insight and absurdity arrive together on a psychedelic — and exactly why what you experience under the drug cannot be accepted as truth on delivery.
The drug wears off in hours, yet the improvements claimed in clinics last weeks. What fills the gap? The strongest candidate right now is structural remodelling.
A 2021 mouse experiment captured it directly. Researchers opened a window in the skull and used two-photon microscopy to re-image the same layer-5 neurons in medial frontal cortex over time, watching their dendritic spines — the little protrusions on a dendrite that serve as physical landing sites for incoming connections, so one more spine is roughly one more usable port. Within 24 hours of a single dose of psilocybin, spine density and spine size were both up by roughly 10% — and were still elevated a month later. Pre-treating with ketanserin blocked the density increase. Same lock again. (How connections strengthen and weaken: Synaptic plasticity; what this cortex does: Prefrontal cortex)
Now back to the loose end from section one. A 2023 finding gave an answer that is almost too tidy: a large share of 5-HT2A receptors aren't on the cell surface at all — they're inside the cell. Serotonin is hydrophilic and can't cross the fatty membrane, so it only ever reaches the pool at the door. Classic psychedelics are lipophilic: they dissolve straight through and turn the inner pool as well — and it's the inner pool that sends the signal to grow spines.
That pushes the field's sharpest question to the front: can you have the plasticity without the hallucination? Following the "lipophilic enough to get in" logic, non-hallucinogenic analogues have already been built, and in animals they still grow spines and still produce antidepressant-like behaviour. If that holds in humans, the claim that mystical experience is a necessary ingredient collapses. There is no decisive human data yet.
The best cross-read for this window isn't an architecture — it's the learning rate, the knob that sets how much a model's parameters are allowed to move on each step. Raise it and the model learns fast and is also far easier to drag off course by whatever batch happens to arrive; a common engineering move is to spike it and then anneal back down, precisely to shake a model out of a bad settled state. The crucial point: the window carries no content. Learning rate governs how much can change, never what it changes into — that is decided entirely by what gets fed in while it's high. Which turns the mystical-sounding clinical phrase set and setting into something quite plain: during those hours, who is present, what you are turning over, and how it gets processed afterwards is the thing actually being written in.
Psilocybin for treatment-resistant depression — depression that hasn't responded to several drugs — is the furthest along. A 233-person phase 2 trial in 2022 found a single 25 mg dose clearly beat the 1 mg dose used as a stand-in placebo at three weeks. A subsequent phase 3 trial met its primary endpoint: at week 6 the 25 mg arm's depression score had fallen 3.6 points further than the comparator, with a highly significant p-value.
But 3.6 points has to be read against a 0–60 scale — statistically solid, clinically modest. Which is why the market shrugged the day the numbers landed. That isn't cynicism; it's the sentence this field most needs stated plainly: "it works" and "it works enough to be worth all this" are two different questions.
A second line routinely conflated with the first is MDMA for PTSD. It isn't a classic psychedelic — it mainly dumps serotonin wholesale rather than working primarily through 5-HT2A. In August 2024 the US FDA declined to approve it for PTSD and asked for another phase 3 trial, citing blinding, safety reporting and durability of effect among its reasons.
And blinding is the field's structural dead end. A placebo control assumes you don't know which arm you're in — with these drugs, almost nobody fails to work it out, and expectancy moves psychiatric endpoints a long way on its own. One clever study nailed this down: 191 people already microdosing were taught to blind themselves (make visually identical capsules and placebos, and use QR codes so the researchers knew the allocation while they didn't). The result: the microdosing group improved, the placebo group improved, and there was no difference between them. The small differences that were detectable are explained by people correctly guessing their arm.
The risks deserve saying out loud too. A family history of psychotic illness is a contraindication — a first episode can be precipitated. A minority are left with persistent visual disturbances (HPPD). And the least-discussed one: chronic repeated low doses keep stimulating 5-HT2B, a route already shown to damage heart valves by withdrawn drugs like fenfluramine and pergolide — which makes "microdose daily for years" precisely the least evidenced and least characterised use of all. Regulation, meanwhile, is loosening: Australia has allowed authorised psychiatrists to prescribe psilocybin and MDMA since 2023, and Oregon and Colorado have built state-level supervised-use systems.
"Loosen the filter and more comes through" is a thought several traditions have had before — and their warnings still hold up: