Doris Tsao and consciousness: what her team’s study really means for AI

An image can disappear from perception without disappearing from brain activity. A preprint from Doris Tsao and colleagues, posted on 16 October 2026, offers a striking account of the timing involved: the recorded neural populations rapidly alternate between representing the perceived stimulus and representing the suppressed one. The dominant percept remains stable.

The study sheds light on how the brain represents content that reaches awareness. It does not explain why subjective experience exists, or show that an AI could become conscious by copying this alternation. Understanding the finding requires keeping three things separate: what reaches the eyes, what neurons encode, and what the animal perceives.

Who is Doris Tsao?

Doris Tsao is a neuroscientist known for her work on how the primate brain encodes faces: understanding how neurons represent facial features makes those representations amenable to decoding. She is now Chief Scientist – Neuro at Astera, having previously been a neuroscience professor at the University of California, Berkeley, and a Howard Hughes Medical Institute (HHMI) Investigator. That background helps place this study in context: it extends research on visual representation rather than establishing a complete theory of consciousness.

How Doris Tsao’s team separates visual input from perception

In everyday vision, those three things usually seem to coincide. A face appears, the brain processes it, and we see it. That makes it difficult to distinguish activity associated with awareness from activity driven simply by the image being present.

A technique called binocular flash suppression pulls them apart. An image is first shown to one eye, then another appears abruptly in front of the other eye. The second image dominates perception, while the first remains physically present but is suppressed from awareness. “Suppressed” does not mean removed from the retina.

Using Neuropixels probes, the researchers simultaneously recorded neural populations in face-selective regions of the inferotemporal and prefrontal cortex. These recordings came from two male macaques. A separate experiment with 14 human participants validated the perceptual behaviour produced by the paradigm. It did not involve recording human neurons.

The human responses therefore support the choice of paradigm, not a claim that these neural results generalise to our species. That would require comparable recordings from human brains.

Roughly 60-millisecond states do not mean consciousness switches off

The key finding emerges from rapid changes in collective neural activity. During periods when both the images and the dominant percept remain stable, the recorded populations switch between a state primarily representing the perceived face and a state primarily representing the suppressed face.

A statistical method called a hidden Markov model estimates median durations of 64 and 59 milliseconds for the two states. The authors describe the pattern as about fifteen switches per second. These are estimates from this dataset, not the discovery of a universal clock governing consciousness.

Crucially, a change in representation is not necessarily a change in experience. The animal does not appear to see the two images flickering back and forth at that rate. The finding is interesting precisely because apparently continuous perception coexists with rapidly changing neural coding.

Averaging activity over a longer interval can blur this temporal structure. Recording many neurons at once lets the researchers investigate whether their states change together, rather than combining responses collected at different times.

Reconstructing a face is not unrestricted mind-reading

To identify what each brief state represents, the team uses a parameterised face space: a set of dimensions describing facial characteristics. Neural activity can then be used to reconstruct features of the perceived face or the suppressed face, depending on the state.

That decoding result is compelling because it makes the signal interpretable. But it depends on a defined stimulus domain and model. It is not a method for freely reconstructing any memory, intention or thought.

Nor is the presence of information about an unseen image entirely new. In a 2020 eLife study, Janis Hesse and Doris Tsao already showed that face cells encoded both perceived and suppressed stimuli. The 2026 contribution concerns their temporal organisation: rapid alternation that averaged responses can obscure.

What the study leaves unanswered about consciousness

The authors explicitly set aside the “hard problem”: why physical activity is accompanied by felt experience at all. Their question is narrower. What distinguishes activity associated with consciously accessible content from activity associated with suppressed content?

The observations reveal an organisation of neural signals, not a complete causal mechanism. They do not establish that imposing this alternation would be sufficient to generate conscious perception, or that removing it would eliminate all experience.

Two animals, selected brain regions and one visual paradigm also cannot stand in for consciousness as a whole. The work remains a preprint: version 1 listed on bioRxiv is not a peer-reviewed publication. Its robustness and generalisability still need testing.

Why conscious AI remains a separate question

An artificial system can maintain several representations, select one, or switch between them. Those functions describe information processing. On their own, they do not tell us whether anything feels like something to that system.

A 2023 report by Butlin and colleagues, including Yoshua Bengio, proposed indicators drawn from scientific theories of consciousness. It suggested that the systems considered at the time were not conscious, while finding no obvious technical barriers to satisfying those indicators. That dated assessment is neither a verdict on every AI system in 2026 nor evidence that meeting the indicators would create subjective experience.

For present-day applications, our business guide to AI agents deals with observable abilities: using tools, retrieving information and completing tasks. Those abilities can be tested without attributing an inner life to the system.

The next step is to ask what evidence would distinguish a system that selects information from one that has an experience. Decoding content, explaining its access to awareness and establishing the existence of subjective experience are different scientific goals. This study sharpens the connection between the first two; it does not collapse all three into one.