A new study out of UC Berkeley and Caltech has produced one of the most striking neuroscience findings in years: the brain doesn’t hold a single, stable representation of what you’re consciously seeing. Instead, in high-level visual and prefrontal cortex, it rapidly toggles between fully representing what you’re aware of and fully representing what you’re not aware of — doing this roughly 15 times every second, even while your actual perception stays constant.
The paper, posted on bioRxiv by Janis Hesse, Frank Lanfranchi, Conor McGrory, Yuelin Shi, Tatiana Engel, and Doris Tsao (whose lab previously cracked the neural code for facial identity), tackles one of the oldest open questions in brain science: when the exact same physical input is sitting in front of you, what makes one part of it reach your awareness while the rest gets filtered out?
The experiment: showing the brain two things at once
The researchers used a classic trick called binocular flash suppression. One image is shown to one eye. A second, different image is then flashed to the other eye. The brain almost always locks onto the new, flashed image as the conscious percept — the first image is still physically present, hitting the retina, but it vanishes from awareness entirely. This setup is powerful because it lets researchers hold the stimulus completely fixed while flipping what the subject actually experiences.
The team first confirmed the illusion works reliably in human volunteers, then recorded from monkeys performing the same task, using Neuropixels probes — high-density electrode arrays capable of listening to hundreds of neurons simultaneously — implanted across four regions: two face-processing areas in inferotemporal cortex (ML and AM) and two regions of prefrontal cortex.
The old puzzle: a “correlate of consciousness” that was always too weak
Earlier studies of binocular rivalry (a cousin of flash suppression) had already shown that neurons respond somewhat differently depending on what’s consciously perceived. But that modulation was frustratingly weak — far weaker than how strongly the same neurons responded when the two images were simply shown one at a time, unambiguously. Worse, the “suppressed,” unseen image could still be partially decoded from brain activity, meaning no one had ever found a clean signal that reflected consciousness and consciousness alone, anywhere in the brain. That gap fed directly into a long-running fight between theories that put the seat of consciousness toward the front of the brain (Global Workspace, Higher-Order theories) versus the back (Integrated Information Theory, Recurrent Processing Theory) — a debate large collaborative projects like COGITATE have tried, and mostly failed, to settle.
What they found instead: the brain is multitasking in time, not space
By presenting 1,500 different faces and using statistical decoding on finely binned neural activity, the researchers discovered that averaging was hiding the real picture. When they looked at 10-millisecond windows instead of averaging across seconds, population activity split cleanly into two alternating states:
- State 1: neurons collectively and purely encode the consciously perceived face. From this activity alone, the researchers could reconstruct a recognizable image of exactly what the animal was seeing.
- State 2: the same population flips to purely encoding the suppressed, unseen face — accurately enough that its features could also be reconstructed into a clear image.
These two states alternate roughly every 60 milliseconds, about 15 cycles per second — an order of magnitude faster than the slow, seconds-long perceptual flips typically studied in rivalry research. Critically, this wasn’t a trick of circular analysis: when the researchers used one random half of neurons to figure out which state the brain was in, and the other, completely independent half to reconstruct the image, the result held up. Individual neurons were shown to actually switch their tuning in lockstep with these transitions, rather than there being one set of cells for “seen” content and another for “unseen” content.
The same alternation, slightly noisier, showed up in prefrontal cortex too — suggesting this isn’t a quirk of one brain area but a broader principle of cortical coding.
This also resolved the old puzzle: the weak “average” signal from decades of rivalry studies was weak because it was an average of two maximally opposite signals — a strong vote for the conscious image and a strong vote for the suppressed one — collapsed together across time. Split them apart in time, and each one turns out to be a near-perfect, high-fidelity signal on its own.
Why this matters for the bigger AI-and-consciousness conversation
This is squarely basic neuroscience, not an AI paper, but it lands in a period where the mechanics of awareness have become unusually relevant to the AI industry — Anthropic has published research on whether language models can introspect on their own internal states, and has also found that large models seem to develop something like a global workspace internally, echoing one of the very theories this paper weighs in on. Executives and researchers have been sparring publicly over whether today’s models show signs of distress in their reasoning traces, whether belief in machine consciousness could eventually curdle into something resembling a new religion, and philosophers like David Chalmers have been pushing increasingly pointed questions about what, if anything, is happening inside these systems when they respond.
What this paper offers that conversation is a concrete, measurable definition of what a “signature of conscious access” actually looks like in a biological brain: not a region, not even a stable pattern, but a specific, fast-switching temporal state that can be told apart from its mirror-image unconscious counterpart with a simple linear decoder. If consciousness really does have a signature this well-defined, it becomes a much more tractable target — something you could, in principle, go looking for, or rule out, in any sufficiently rich information-processing system, biological or artificial.
What’s left unanswered
The paper is explicit about what it hasn’t explained: why some of these 60-millisecond windows get “selected” for awareness and others don’t. The authors speculate that the suppressed micro-epochs might let the brain quietly test out alternative interpretations of a scene in the background, while the conscious epochs represent whichever interpretation currently gets routed onward to circuits involved in memory, decision-making, and reporting. Figuring out what actually does that routing — and why — is, in their words, the next problem to crack.