[Written by Claude and ChatGPT. Image by ChatGPT]
Every night, without a script or a director, your brain generates an entire world — places you’ve never been, people who don’t quite exist, conversations that make perfect sense until you wake up. And somehow, you’re usually in it: not always steering events, but embedded in them, seeing, feeling, reacting, accepting impossible logic as completely normal.
The classic explanation is activation-synthesis theory: random signals during REM sleep get “synthesized” into a coherent experience. That idea is historically important, but it’s really too simple. A more accurate picture: the brain temporarily builds a world out of its own memories, predictions, emotions, and body models — with a version of you embedded inside it.
A brain that’s gone inward
During REM sleep, brainstem circuits push the cortex into an unusually activated state. In animal studies, researchers have identified characteristic bursts called PGO waves — ponto-geniculo-occipital activity, tracing a path from the pons through the thalamus to the visual cortex. Human researchers can’t observe this as directly, so most of what we know comes from animal work and indirect measures like EEG and fMRI.
What’s clearer is the chemistry: acetylcholine rises, while serotonin and norepinephrine — the systems behind focused attention and impulse control — drop off substantially. Meanwhile, most major skeletal muscles are inhibited, so the actions your brain generates never get carried out.
The result is a brain that’s highly active but almost entirely cut off from the outside world, running under a very different chemical regime than waking life. That’s the setup that lets internally generated experience take over.
Where the images actually come from
Here’s where “random signals become pictures” breaks down. The visual cortex isn’t a blank screen waiting for the brainstem to project something onto it — even while awake, your brain is constantly filling gaps and resolving ambiguity using memory and expectation, not just raw sensory data.
With external input mostly shut off during REM, the balance flips: instead of world → senses → brain → experience, it’s closer to brain → internal activity → simulated world → experience. PGO waves don’t need to carry little encoded pictures for this to work — there’s no real evidence for that kind of one-to-one mechanism. They more plausibly just help drive the brain into the activated state; the actual content comes from an enormous existing library of faces, rooms, fears, and relationships the brain doesn’t have to invent from nothing. It just has to recombine what it already has.
Why faces stay faces, and anatomy stays correct
This raises a real puzzle: if dream imagery is generated with no sensory input to check it against, why don’t faces come out warped, or animals with limbs in the wrong places? The answer is that structural correctness was never something your eyes were maintaining in the first place — it’s stored as compressed structural knowledge, and dreaming draws on that store directly.
A face isn’t held in the brain as a picture; regions like the fusiform face area encode something closer to a generative template — the relationships between eyes, nose, and mouth, the proportions and symmetry that make something face-shaped, abstracted from a lifetime of examples. When dream activity recruits this circuitry, it instantiates the template, which comes with correct structure built in by default. The same logic covers anatomy generally: you’ve never consciously catalogued “a dog has four legs positioned just so,” but visual and category-representation areas have encoded that structure through massive exposure, the same way they encode “chairs have legs and a seat” without you memorizing it as a fact. Generating a dream face or a dream dog means pulling from a template that already has the right structure baked in — not painting something freehand and hoping the parts land correctly.
This is also why dream weirdness shows up somewhere else instead. Notice where dreams actually go wrong — it’s rarely “the eyes are misplaced.” It’s context, continuity, and identity: the wrong person’s face on a familiar body, a room that shouldn’t connect to another room, someone long dead casually present. That’s exactly the pattern you’d expect if low-level structural templates stay intact while the higher-level systems responsible for binding those templates to the right context, timeline, and identity are the ones running unchecked. The parts are correct; the assembly is what gets scrambled. It’s a useful contrast with early AI image generators, which famously botched hands and teeth because they lacked a robust structural model of anatomy underneath — your brain has no such weakness for its core categories, and dreams inherit that robustness for free.
The same question scales up to whole scenes, though: a single correct face is one thing, but a coherent room — consistent lighting, correct spatial layout, several people who all look right at once — is a bigger binding problem. This turns out not to be a separate mystery either, because scene coherence isn’t generated fresh each dream — it’s handled by a system that does this exact job constantly, even while you’re awake. Right now, without looking, you can picture your kitchen: its layout, where the counter sits relative to the door, roughly how it’s lit. You’re not consulting a stored photograph; you’re reconstructing a spatial scene from structural knowledge, using what researchers call scene construction — a network involving the hippocampus, parahippocampal place area, and retrosplenial cortex that’s active whenever you imagine a place or mentally rehearse a scenario. Dreaming appears to recruit this same machinery, so a coherent dream scene isn’t a special nighttime trick — it’s the identical process behind daydreaming, just running with less oversight and no off switch.
Crucially, this kind of coherence just means internal consistency, not verification against reality. The system doesn’t need external input to check “is this room correct” — it only has to maintain one consistent spatial frame and populate it, which is a fundamentally easier job than independently generating thousands of details that all happen to agree. Once the frame is set, everything slotted into it inherits that consistency automatically. And it’s less flawless than it feels in the moment — spatial impossibilities and subtly wrong faces are common in dreams, but the systems that would normally flag “this violates physical consistency” are exactly the ones suppressed during REM, so the glitches don’t register as errors, and memory tends to smooth them into something that sounds more coherent than what was actually experienced.
Memory supplies the raw material — but doesn’t just replay
Dreams constantly draw on waking life, sometimes directly (you dream about someone you saw that day) and sometimes almost unrecognizably transformed (your childhood home, rearranged; an old friend playing a coworker’s role). Sleep is deeply involved in memory processing, and there’s real evidence that memory-related neural activity gets reactivated during sleep.
But dreaming isn’t playback — it’s reconstruction. Recent and old memories collide, emotional tags attach to unrelated settings, familiar faces show up in unfamiliar roles. That recombination is inherently creative, which is part of why dream logic feels associative rather than causal.
So why do dreams feel like they have a plot?
If dreams are built from recombined fragments, why don’t they feel like a slideshow of unrelated images? A few things create continuity, or at least the strong impression of it:
- Associative momentum. Activating one node — “high school,” say — pulls in a whole connected network: classrooms, old anxieties, specific people. The dream doesn’t follow logical causality so much as it follows association, one thing pulling the next along with it.
- Emotional continuity. Settings and characters can shift completely while an underlying feeling — dread, longing, urgency — stays constant, giving disconnected scenes a shared thread.
- The brain’s hunger for coherence. Human cognition is exceptionally good at building explanations from incomplete information, a tendency well documented in research on perception, memory, and confabulation. Dreaming seems to exploit this same drive: rather than flagging a discontinuity (your teacher is suddenly your mother), the dreaming brain just accepts it and moves on — until you wake up.
- Memory selection. Most dreams vanish almost instantly. What survives into your morning memory is probably not a random sample — it skews toward the more vivid, emotional, or coherent dreams, which is part of why dreaming feels more narratively organized in retrospect than it may actually be in the moment.
Why are you in almost every dream?
The strangest part of dreaming isn’t the bizarre content — it’s that you’re reliably there experiencing it. One likely reason: the brain isn’t just building a scene, it’s building a model of a self within the scene. Waking experience is already organized around an embodied point of view — “I am here, this is my body, that happened to me” — and your brain maintains this model continuously. It doesn’t switch off at sleep onset; it becomes part of the internally generated world instead. The brain already has a richly detailed model of you running at all times, so it doesn’t need to invent a protagonist — it just keeps using the one it has.
That said, the dream-self isn’t fixed. Sometimes you watch yourself from outside, inhabit a different body, or become someone else entirely without questioning it. This suggests the sense of self in a dream isn’t one switch flipped on nightly, but something assembled dynamically from several separate systems — body representation, perspective, memory, agency — that dreaming can recombine or destabilize just like everything else.
Actor, or audience?
This is also why “I’m in the dream” doesn’t automatically mean “I’m driving the dream.” Being the first-person locus a dream is organized around — the “you” everything else is arranged relative to — is a different thing from having control over what happens. You can be centrally embedded in a scene, feeling it happen to you, startled or moved by it, while having essentially no influence over the plot. That’s compatible with dreaming feeling like watching a movie — just a movie you’re inside of, rather than one directed from a seat outside it.
Dreams vary a lot in how agentic they feel, and this tracks real neural variability. Some dreams involve deliberately choosing, running, deciding to speak — engaging the same motor-planning and decision circuitry active in waking life; brain systems involved in movement planning stay active during REM even though the muscles that would execute those movements are inhibited, which is why running in a dream can feel physically real without your legs ever moving. Other dreams are much closer to pure spectatorship — events unfold and you’re carried along, unable to influence outcomes even when you try, the classic “trying to run but can’t” dream. This is thought to relate to how much prefrontal and self-monitoring activity happens to be online at a given moment, since REM isn’t a uniform state — it fluctuates in depth and in which regions are more or less suppressed from one stretch of the dream to the next. Occasionally that monitoring activity comes back online enough that you realize you’re dreaming and gain unusual control over the scene — lucid dreaming — though the precise mechanics of that shift are still being worked out.
If your own dreams tend to feel more like watching than acting, that’s a legitimate and common point on this spectrum, not a contradiction of the underlying mechanism — it likely just means the “experiencer” side of your dream-self is more consistently active than the “decider” side.
What PGO waves are actually doing — and what we don’t know
It’s worth being precise here about the limits of the evidence. PGO activity is strongly tied to REM sleep and likely helps produce and sustain that state; it may also interact with memory processing and sensory activation. There are real hypotheses that it contributes to the visual content of dreams, largely based on animal research.
But there’s no solid evidence for a clean chain like PGO wave → image → memory → dream scene. The truth is almost certainly messier, and because so much of the detailed work is done in animals, with very little invasive recording possible in humans, real uncertainty remains about how directly these signals shape dream content. That’s not a weakness in the theory — it’s an honest reflection of where the science currently stands.
So what is a dream?
The simplest description: an internally generated conscious world, built with no single “dream center” in charge. It emerges from brainstem REM circuitry, sensory and associative cortex, memory systems, emotional networks, and the machinery that constructs a sense of self and agency — all running at once, generating the setting, the characters, the emotional tone, and the experiencer simultaneously, while structural templates for faces, bodies, and spaces quietly keep the form of the scene intact even as its content runs loose.
Because the experiencer is part of the simulation, you don’t perceive it as a simulation. You perceive it as reality — sometimes as a place you’re moving through, sometimes as one you’re merely caught inside of — until you wake up, the world disappears, and what’s left is sometimes just a feeling, a face, or the strange certainty that you were somewhere else entirely.
