Beyond the Surface: A Neurocognitive Perspective on Deepfake Face Perception
Quinn Tatiyamaneekul
Think about the last time a face online made you stop scrolling. Not because you recognized it. Because you almost did.
At first glance, everything seems fine. The proportions look right, nothing is obviously wrong. But something feels off before you can name what it is. That hesitation matters more than it seems. It suggests the brain is not just identifying faces but continuously auditing them, running a background check on incoming information before conscious awareness even enters the picture. Which raises a genuinely strange question: how does the brain decide what is real?
Deepfake technology has become a surprisingly useful tool for studying exactly this. These AI-generated faces are engineered to pass as human, and they often do, but they occupy an uncomfortable middle ground between authentic and artificial. That gap is scientifically valuable. By testing what happens when something looks realistic but is not genuine, researchers can probe the brain's verification process directly. No single field explains it fully. Neuroscience identifies the mechanism, psychology explains why it so often fails to surface as conscious knowledge, and computer science provides both the stimulus and the evidence for what is hardest to fake. Read together, they reveal something striking: facial recognition is not a passive identification system. It is an active one built around trust.
The neuroscience findings are, frankly, more dramatic than you might expect. A 2023 EEG study by Tarchi and colleagues found measurably different patterns of brain activation depending on whether participants viewed real or synthetic faces (PMC, 2023). Real faces produced right-hemisphere dominance across multiple frequency bands, while deepfakes primarily recruited frontal and occipital regions (PMC, 2023). More tellingly, a component called the N250, which tracks facial identity processing, peaked significantly faster for authentic faces (PMC, 2023). The brain was not just recognizing faces differently. It was processing real ones with noticeably greater fluency, as though it found them easier to believe.
A 2024 study by Chen and colleagues added something unexpected. Using steady-state visual evoked potentials, they found that neural response strength did not increase linearly with realism. Both highly artificial faces and fully photographic ones triggered stronger responses than near-human intermediate ones (Nature, 2024). The brain does not simply relax as faces become more realistic. It treats the almost-real as a special category, one that generates its own distinct neural signature. This matters because it suggests realism is not a passive spectrum the brain slides along. It is something the brain actively encodes, with near-misses setting off their own alarm (Nature, 2024).
The N400 finding, though, is the one that genuinely surprised me when I read it. This neural signal, typically associated with expectation violations, fires when a sentence ends with the wrong word or a face does not quite cohere as a real one. A 2026 study by Becker and colleagues found that deepfake faces elicited significantly higher N400 amplitudes than real footage, even when participants reported no awareness of any manipulation at all (ScienceDirect, 2026). The brain flagged the mismatch before the person knew to look for one (ScienceDirect, 2026). Awareness, when it arrives, arrives late.
This is where psychology fills in what neuroscience cannot explain on its own. The uncanny valley describes the unease people report when a face closely resembles a human but falls just short of convincing. Schindler and colleagues traced this discomfort to specific perceptual mismatches: gaze that lingers fractionally too long, micro-expressions arriving a beat out of phase, emotional responses that do not quite match context (PMC, 2017). These are exactly the dynamic cues current deepfake technology struggles most to replicate. The weird thing, though, is that knowing this does not make people better at detecting fakes consciously. A 2026 study by Pehlivanoglu and colleagues found that while average discrimination performance was above chance, variation across individuals was enormous, with some performing near randomly (Springer, 2025). People's ratings of real versus fake faces showed systematic biases that did not track their actual accuracy (Springer, 2025). The gap between what the brain registers implicitly and what people can consciously report is not random noise. It reflects something fundamental about how perception is organized: the implicit processing runs first, and what reaches awareness is already a compressed summary, not the raw signal.
Computer science sits in an unusual position here. Deepfakes are not only the subject of the research but the tool that makes the research possible. What AI can replicate with high accuracy is static: symmetry, skin texture, lighting. What it still struggles with is temporal: the subtle timing of a blink, the micro-muscular choreography of a genuine smile, the way real emotional expressions unfold over fractions of a second. Frontal delta activity, Becker and colleagues note, appears specifically linked to perceiving naturalistic facial motion (ScienceDirect, 2026). This means the limits of the technology map almost perfectly onto the features the brain is most sensitive to. Studying what AI cannot fake reveals what human perception most depends on.
An evolutionary framing rounds it out. Humans spent a long time in environments where reading faces accurately carried real stakes. Detecting subtle inconsistencies in gaze or expression timing could signal deception. The brain's implicit response to deepfakes is not a malfunction. It is an old detection system encountering a stimulus it was never designed to encounter: a face that passes every static test and fails the dynamic ones.
The stakes of understanding this go well beyond the laboratory. Deepfakes are already being used to fabricate political speeches, generate non-consensual intimate imagery, and erode the evidentiary value of video in legal proceedings. A society where footage can no longer be assumed authentic has serious implications for journalism, accountability, and individual harm. The brain's implicit detection system, however finely tuned, was not built to operate at the scale of a social media feed where a fake can reach millions before anyone consciously evaluates it.
What deepfakes reveal, when examined across all three fields, is that perception is not a camera. It is a judgment. One the brain begins rendering before you even know the question is being asked.
About the Author
Quinn Tatiyamaneekul is a student at the Bangkok Patana School.
References
At first glance, everything seems fine. The proportions look right, nothing is obviously wrong. But something feels off before you can name what it is. That hesitation matters more than it seems. It suggests the brain is not just identifying faces but continuously auditing them, running a background check on incoming information before conscious awareness even enters the picture. Which raises a genuinely strange question: how does the brain decide what is real?
Deepfake technology has become a surprisingly useful tool for studying exactly this. These AI-generated faces are engineered to pass as human, and they often do, but they occupy an uncomfortable middle ground between authentic and artificial. That gap is scientifically valuable. By testing what happens when something looks realistic but is not genuine, researchers can probe the brain's verification process directly. No single field explains it fully. Neuroscience identifies the mechanism, psychology explains why it so often fails to surface as conscious knowledge, and computer science provides both the stimulus and the evidence for what is hardest to fake. Read together, they reveal something striking: facial recognition is not a passive identification system. It is an active one built around trust.
The neuroscience findings are, frankly, more dramatic than you might expect. A 2023 EEG study by Tarchi and colleagues found measurably different patterns of brain activation depending on whether participants viewed real or synthetic faces (PMC, 2023). Real faces produced right-hemisphere dominance across multiple frequency bands, while deepfakes primarily recruited frontal and occipital regions (PMC, 2023). More tellingly, a component called the N250, which tracks facial identity processing, peaked significantly faster for authentic faces (PMC, 2023). The brain was not just recognizing faces differently. It was processing real ones with noticeably greater fluency, as though it found them easier to believe.
A 2024 study by Chen and colleagues added something unexpected. Using steady-state visual evoked potentials, they found that neural response strength did not increase linearly with realism. Both highly artificial faces and fully photographic ones triggered stronger responses than near-human intermediate ones (Nature, 2024). The brain does not simply relax as faces become more realistic. It treats the almost-real as a special category, one that generates its own distinct neural signature. This matters because it suggests realism is not a passive spectrum the brain slides along. It is something the brain actively encodes, with near-misses setting off their own alarm (Nature, 2024).
The N400 finding, though, is the one that genuinely surprised me when I read it. This neural signal, typically associated with expectation violations, fires when a sentence ends with the wrong word or a face does not quite cohere as a real one. A 2026 study by Becker and colleagues found that deepfake faces elicited significantly higher N400 amplitudes than real footage, even when participants reported no awareness of any manipulation at all (ScienceDirect, 2026). The brain flagged the mismatch before the person knew to look for one (ScienceDirect, 2026). Awareness, when it arrives, arrives late.
This is where psychology fills in what neuroscience cannot explain on its own. The uncanny valley describes the unease people report when a face closely resembles a human but falls just short of convincing. Schindler and colleagues traced this discomfort to specific perceptual mismatches: gaze that lingers fractionally too long, micro-expressions arriving a beat out of phase, emotional responses that do not quite match context (PMC, 2017). These are exactly the dynamic cues current deepfake technology struggles most to replicate. The weird thing, though, is that knowing this does not make people better at detecting fakes consciously. A 2026 study by Pehlivanoglu and colleagues found that while average discrimination performance was above chance, variation across individuals was enormous, with some performing near randomly (Springer, 2025). People's ratings of real versus fake faces showed systematic biases that did not track their actual accuracy (Springer, 2025). The gap between what the brain registers implicitly and what people can consciously report is not random noise. It reflects something fundamental about how perception is organized: the implicit processing runs first, and what reaches awareness is already a compressed summary, not the raw signal.
Computer science sits in an unusual position here. Deepfakes are not only the subject of the research but the tool that makes the research possible. What AI can replicate with high accuracy is static: symmetry, skin texture, lighting. What it still struggles with is temporal: the subtle timing of a blink, the micro-muscular choreography of a genuine smile, the way real emotional expressions unfold over fractions of a second. Frontal delta activity, Becker and colleagues note, appears specifically linked to perceiving naturalistic facial motion (ScienceDirect, 2026). This means the limits of the technology map almost perfectly onto the features the brain is most sensitive to. Studying what AI cannot fake reveals what human perception most depends on.
An evolutionary framing rounds it out. Humans spent a long time in environments where reading faces accurately carried real stakes. Detecting subtle inconsistencies in gaze or expression timing could signal deception. The brain's implicit response to deepfakes is not a malfunction. It is an old detection system encountering a stimulus it was never designed to encounter: a face that passes every static test and fails the dynamic ones.
The stakes of understanding this go well beyond the laboratory. Deepfakes are already being used to fabricate political speeches, generate non-consensual intimate imagery, and erode the evidentiary value of video in legal proceedings. A society where footage can no longer be assumed authentic has serious implications for journalism, accountability, and individual harm. The brain's implicit detection system, however finely tuned, was not built to operate at the scale of a social media feed where a fake can reach millions before anyone consciously evaluates it.
What deepfakes reveal, when examined across all three fields, is that perception is not a camera. It is a judgment. One the brain begins rendering before you even know the question is being asked.
About the Author
Quinn Tatiyamaneekul is a student at the Bangkok Patana School.
References
- Becker, C., Conduit, R., Chouinard, P.A. and Laycock, R. (2026). Brain Responses to Deepfakes and Real Videos of Emotional Facial Expressions Reveal Detection Without Awareness. Computers in Human Behavior, p.108923. doi:https://doi.org/10.1016/j.chb.2026.108923.
- Chen, Y., Stephani, T., Bagdasarian, M.T., Hilsmann, A., Eisert, P., Arno Villringer, Bosse, S., Gaebler, M. and Nikulin, V.V. (2024). Realness of face images can be decoded from non-linear modulation of EEG responses. Scientific Reports, [online] 14(1). doi:https://doi.org/10.1038/s41598-024-56130-1.
- Didem Pehlivanoglu, Zhu, M., Zhen, J., Gagnon-Roberge, A.A., Kern, R.K., Woodard, D., Cahill, B.S. and Ebner, N.C. (2026). Is this real? Susceptibility to deepfakes in machines and humans. Cognitive Research Principles and Implications, 11(1), pp.3–3. doi:https://doi.org/10.1186/s41235-025-00700-y.
- Pietro Tarchi, Maria Chiara Lanini, Frassineti, L. and Lanatà, A. (2023). Real and Deepfake Face Recognition: An EEG Study on Cognitive and Emotive Implications. Brain Sciences, 13(9), pp.1233–1233. doi:https://doi.org/10.3390/brainsci13091233.
- Schindler, S., Zell, E., Botsch, M. and Kissler, J. (2017). Differential Effects of face-realism and Emotion on event-related Brain Potentials and Their Implications for the Uncanny Valley Theory. Scientific Reports, 7(1). doi:https://doi.org/10.1038/srep45003.