There is no metronome in your head. The beat is what "stop" sounds like when everyone hears it at once — and the brain uses it to decide which line is open right now.
When everyone in a room talks at once you can't make out a word; let them take turns and the same room and the same pair of ears catch everything. The brain faces an extreme version of that problem: tens of billions of cells wired densely together, countless lines shouting at any instant, and no way to run a private wire for every pair of cells. So it does something else — it slices up time. Put an electrode on any piece of living brain tissue and you will see voltage rising and falling rhythmically. This issue is about that rise and fall: where it comes from, what the brain does with it, and an argument that is still unsettled — is it the machinery doing the work, or just engine noise?
First, what is actually being measured. Electrodes on the scalp (EEG) and electrodes pushed into tissue do not pick up individual cells firing — that's too fast and too scattered, and it drowns out within a short distance. What they pick up is the sum of membrane currents across thousands of cells. For those currents to add up rather than cancel, one demanding condition has to hold: the cells must be geometrically aligned and active at the same time. Cortical pyramidal cells happen to stand in parallel like a field of saplings, which is the only reason their currents can build into a signal detectable through the skull (Neurons & action potentials). Hidden in that is a conclusion people routinely get backwards: a strong brain wave doesn't mean more firing, only firing in step.
So where does "in step" come from? The cheap way is not to issue orders one cell at a time, but to interrupt everyone at once. The cortex has a class of cells devoted to exactly that: they send nothing to distant regions and just pull a local switch, one spike pushing down a whole neighbourhood (Inhibitory interneurons). That sets a loop spinning on its own: a group of excitatory cells fires, which recruits the inhibitory cells; the inhibitory cells clamp the whole patch, so everyone goes silent together; inhibition decays; the pent-up excitatory cells fire together again. How fast inhibition wears off sets how fast the loop turns — about twenty-odd milliseconds per round, which works out to roughly forty times a second: the band we call gamma.
None of this is armchair reasoning. In 2009 two groups used optogenetics — fitting one specific cell type with a light-sensitive switch, then using blue light to dictate exactly when those cells fire — to test this very claim: drive the fast-spiking inhibitory cells alone and the cortex immediately produces clean gamma; drive the excitatory cells instead and it doesn't (Cardin et al.; Sohal et al.). The beat is kept by inhibition.
Once there is a cycle, something comes for free: phase — where you are within one rise and fall, whether it's just climbing, at the crest, or already dropping. "This cell fired once" means two different things depending on the phase it lands on. Alongside how fast a cell fires, the brain gets a nearly free second channel: when it fires.
The prettiest evidence is in the rat hippocampus. It holds a class of cells that fire only when the animal reaches one particular spot (Hippocampus & entorhinal cortex). In 1993 O'Keefe and Recce found that as a rat crosses that spot from one end to the other, the cell's spikes arrive at an earlier phase on every successive cycle — late in the slow wave on the way in, creeping forward the deeper it goes. So "how far through this place has the rat got" is told more precisely by when the spikes come than by how many there are. This is phase precession, and it's the hardest brick in the claim that phase genuinely carries information.
Phase can also serve as a set of slots. Each cycle of the slow wave (theta, four to eight per second; slower and more intermittent in humans) has room for seven or eight fast gamma cycles inside it. Lisman's reading: one item per fast cycle, with a whole slow cycle replaying the sequence in order — holding several things at once rests not on several slots but on several moments. That count lines up with working-memory capacity (the "about seven" from Topic 3), which proves nothing but is hard to write off as coincidence. The same nesting of fast waves riding on slow ones has been measured in human cortex (Canolty et al. 2006).
Today's language models (the ChatGPT family) have an unavoidable problem: they swallow a whole sentence in parallel, all at once, and on their own can't tell "cat chases dog" from "dog chases cat." The standard fix is positional encoding — a bank of sine waves at different rates computes a string of numbers as each position's fingerprint, and the model reads order off that string. It is the same trick the brain uses with phase: turn "which one came first" into a location on a wave, so order becomes a quantity you can compute with directly. The difference is worth keeping in mind too: the model's version is a fixed lookup table of coordinates, while the brain's oscillations are alive — their speed and alignment shift from moment to moment.
Now put the two together. If inhibition is clamping the whole patch cycle after cycle, then the receiving side has, within every cycle, a stretch where it is unusually receptive and a stretch where it is essentially deaf. So the same message means completely different things depending on when it arrives: catch the open window and a few inputs summing together push the cell past threshold; hit the closed one and identical inputs get eaten by inhibition and nothing happens.
From this Fries built a strong claim, communication through coherence: whether two regions have an anatomical connection only settles what can happen; what settles whether the line is open right now is whether the two rhythms are in phase. One and the same wire can be switched on and off by adjusting phase — so the brain never needs a dedicated cable for every possible pairing.
Supporting observations are plentiful. When a monkey attends to something, the patch of cells processing that thing shows markedly stronger gamma synchrony while the patch processing a nearby distracter does not (Fries et al. 2001) — attention here looks less like turning up the volume and more like setting the clocks to match. There's a division of labour by direction too: influences travelling upstream (feedforward) ride gamma while those coming back down (feedback) ride the slower beta, each direction on its own band (Bastos et al. 2015), so the two don't collide on the same wire (Visual pathway). And there is a mirror-image case: alpha (eight to twelve per second) actually grows stronger over the side you intend to ignore — as if lengthening that channel's closed window, deliberately tuning it to a rhythm that is hard to be heard on (Thalamus).
This is one of the rare places where brains and artificial networks fail to line up — and mismatches often carry more information than matches. In an artificial network, who talks to whom is written entirely in the weights and the attention scores; a layer is one parallel shot of arithmetic with no "when" dimension at all, and it can afford a separate set of parameters for every pairing. The brain has no such luxury: the hardware is a fixed set of wires, so it has to multiplex the same wiring in time, switching who is listening to whom by matching phase. Which is precisely what spiking and neuromorphic chips are trying to bring back (Topic 40).
That went down a little too smoothly; time to stop. One camp has a weighty objection: oscillations may be largely a by-product. Ray and Maunsell point to a solid difficulty — gamma's frequency isn't fixed, it drifts from the low thirties up to eighty hertz with the contrast of what's on screen. If two areas are looking at scenes of different contrast, their rhythms don't match to begin with, so how could phase-matching carry the traffic? The more general trap: two regions can be synchronous simply because they receive a common input, not because they are talking to each other — correlation isn't causation (Topic 34 is devoted to this class of mistake).
But on the other side sit several cases where you can reach in, change something, and watch the outcome move.
During sleep. Deep sleep runs a three-layer nested rhythm: a slow oscillation of roughly one cycle per second pushes the whole cortex back and forth between active and silent; on each active stretch a spindle of a dozen or so hertz grows; and inside the spindle sit brief hippocampal ripples around a hundred hertz, in which the day's experience is replayed (Topics 22 and 32 cover this system).
The point is that you can intervene. Sleep with headphones on while a machine tracks the slow oscillation live and plays a soft click exactly at the up-phase, and next morning's word recall improves; move the identical click to the opposite phase and the benefit disappears (Ngo et al. 2013). Same sound, same loudness, same number of repetitions — only the phase differs. That is about as clean a demonstration of phase having causal force as you can ask for.
When things go wrong. An epileptic seizure is synchrony out of control: a whole population that ought to be staggered gets dragged onto one beat. In Parkinson's disease, excessive beta in the basal ganglia tracks the difficulty in getting movements started; in February 2025 the US FDA approved the first "adaptive" deep brain stimulation system — electrodes that listen to how strong beta is and adjust the stimulation in real time. Here the rhythm is no longer a theory; it is a signal being read out and used for control.
But don't buy the lamp yet. Flickering light and clicks at 40 Hz look striking in mouse models of Alzheimer's (Iaccarino et al. 2016), yet human trials remain small and inconsistent — nowhere near the point of recommendation.
So how should the argument be settled? Roughly like this: oscillations really are a by-product of everyone being clamped together, and they really have been put to use — the two are not in tension at all (a heartbeat is a by-product of pumping blood, and doctors read it as a signal anyway). The question worth asking was never the grand "do oscillations do anything," but "in this specific link, has anyone actually perturbed it, and did the outcome change?"
"The cheapest way to make many parts act together is to give them a shared beat" — several fields far from neuroscience each found their way to this one: