Wake someone out of the deepest, slowest, most switched-off-looking stage of sleep and there's roughly a coin flip's chance they'll tell you they were experiencing something. That number comes out of the largest pooled dream dataset anyone has assembled, and it does not fit the story most of us were taught about when dreaming happens.
The dataset is called DREAM, short for Dream EEG and Mentation, and it landed in Nature Communications on August 13, 2025. Co-first authors William Wong at Monash University and Rubén Herzog at the Paris Brain Institute, with Katja Valli at the University of Turku and Naotsugu Tsuchiya at Monash as corresponding authors, pulled 20 separate sleep studies into one standardized pile. The initial release holds 505 participants and 2,643 awakenings. Every awakening pairs the brain recording running up to the moment of waking with a classified report of what the sleeper said they had been experiencing.
That harmonizing step is the whole point. Individual dream labs run tiny: a dozen participants, a few dozen awakenings, each lab using its own wording for the wake-up question and its own rules for what counts as a dream. Pooling them has been close to impossible. Thirty-seven institutions agreed to a common format so the numbers could finally be stacked.
This initiative is unprecedented. Experimental studies on dreams are very costly, and data sharing remains rare in this field.Naotsugu Tsuchiya, Monash University
The stage you wake from changes the odds, not the answer
Each awakening got sorted into one of three buckets. Experience means the sleeper reported something and could say what it was. Experience without recall means they were certain something had been going on but the content was gone, the state you may know from the maddening blank morning. No experience means nothing at all, the sleeper's own sense that the last stretch of time was simply absent.
Here is how that broke down. Out of 515 REM awakenings, 81% produced a full report with recalled content. Out of 861 awakenings from N2, the ordinary mid-depth sleep where you spend most of the night, 56% did. Out of 64 awakenings from N3, the deep slow-wave stage, 48% did. And from N1, the thin drifting layer at sleep onset, 88% of 110 awakenings produced a report, the highest rate of any stage in the set.
Read those in order and the tidy REM-equals-dreaming equation falls apart. REM is the best single bet, sure. But a sleeper hauled out of slow-wave sleep tells you about an experience nearly half the time. That is not noise, and it is not people confabulating to please the researcher. The paper's own framing puts NREM report rates somewhere in the 40 to 60% band across the literature. Dreaming runs across the whole night. It simply concentrates in REM.
One honest caveat, because the number that will get quoted loudest is the shakiest one: that N3 figure rests on 64 awakenings. Sixty-four. Deep sleep is brutally hard to sample because participants keep sliding out of it and because waking someone from N3 produces a groggy, half-verbal person who's a poor witness to their own night. Treat 48% as a real signal with wide uncertainty around it, not a settled constant.
The doorway almost nobody uses
The finding I'd actually change my behavior over is N1. Sleep onset, the first minute or two after you tip over the edge, returned experience reports 88% of the time. More often than REM.
Anyone who's ever jolted awake on a couch with a fragment of nonsense still playing knows this state. It's hypnagogia: the drifting faces, the falling sensation, the sentence you were sure made sense a second ago. Sleep researchers have known it produces imagery for decades, but it gets treated as a curiosity because the episodes are short and structurally simple compared with a full REM narrative. The DREAM numbers say that if your goal is to catch conscious content coming off your own sleeping brain, the moment you fall asleep is the most reliable window you have, and it's the one you cross unattended every single night.
How to use the sleep-onset window
Take a short afternoon nap with an alarm set for ten to fifteen minutes. You'll cross into N1 within a few minutes and get pulled back out before N2 swallows the memory. Write down whatever's in your head the instant you open your eyes, even if it's one image or a fragment of a sentence with no story attached. Do this a handful of times and you'll build a feel for the texture of your own hypnagogia, which is the same recognition skill that the WBTB method trains at the other end of the night. It's also the raw material for wake-initiated lucid dreaming, where you ride that threshold consciously instead of blacking out across it.
The electrodes still can't tell
The team also tried the obvious next thing: predict from the EEG alone whether the sleeper would report an experience. They got above chance, and not by much. Peak performance was an AUC of 0.586 in NREM and 0.700 in REM.
AUC is worth understanding because it's the number that separates a real finding from a product claim. A score of 0.5 is a coin flip. A score of 1.0 is a perfect detector. So 0.586 in NREM means the brainwaves carry a faint, statistically real trace of whether someone is having an experience, and almost nothing usable. Even the REM figure of 0.700, respectable for this kind of work, would be an unacceptable diagnostic anywhere in medicine.
Keep that number in your pocket the next time a wearable's marketing page starts talking about dream tracking. Any headband or ring claiming to know when you're dreaming, and to time a cue or an alarm to it, is running on a stage estimate, not a dream detection. The best pooled attempt by the people with the actual data can barely beat a coin. Stage-based timing still works for practical purposes, because REM is where the odds concentrate, but the device isn't seeing your dream. It's playing the base rate.
Why the REM story stuck anyway
The equation dates to the 1950s, when researchers first noticed that waking people during periods of rapid eye movement produced vivid, detailed reports far more often than waking them at other times. That was a genuine discovery and it built modern sleep science. The trouble is what happened next: sleep stages got formalized as scoring categories based on electrical patterns and muscle tone, and dreaming quietly got filed as a property of one of those categories.
But the stages were never designed to measure consciousness. They were designed to describe what the electrodes do. Sorting a night into N1, N2, N3, and REM is a bookkeeping convention that turned out to be useful, and treating it as a map of when the lights are on inside was always a leap. The DREAM database is the first dataset large enough to show the seam clearly, which is also why what the brain does during a lucid dream has been so hard to pin to any single stage signature.
If you keep a journal, the practical takeaway is small and real: stop assuming a non-REM waking is a wasted one. The odds are lower outside REM, not zero, and at sleep onset they're better than anywhere. The stage tells you what to bet. It was never going to tell you what happened.