Conscious and unconscious eye contact at the limits of vision
nvidia/Gemma-4-26B-A4B-NVFP4 · academic_accessible/eval 92%/5 min read/Jul 14, 2026
Generated by a local model (nvidia/Gemma-4-26B-A4B-NVFP4) from a scientific paper, claim-checked against the full text. Provenance is open by design.
Eye Contact Enhances Face Detection at the Limits of Vision Before Conscious Awareness Emerges
When you catch someone staring at you, your brain reacts almost instantly. This social reflex is one of the most powerful signals in human life. It helps us regulate attention and social hierarchy. Researchers have long suspected that our brains prioritize eye contact. But a fundamental question remains: exactly how much visual information is required to register a gaze? And does that registration happen before we are even aware of it?
Current research into social perception often relies on "suppression" techniques. These methods intentionally hide information from one eye to see how it influences the other. While these methods suggest that direct gaze gains access to awareness more easily, they are difficult to interpret. They cannot easily distinguish whether the brain is truly detecting a signal from minimal input. Or if the experimental method itself is biasing the results. We lack a clear understanding of the precise "threshold" where a fleeting glance transforms from raw sensory data into a conscious social signal.
A new study from the Université libre de Bruxelles and the Karolinska Institutet proposes that the visual system detects eye contact far earlier than we realize. The authors report that direct gaze allows the brain to localize a face even with incredibly brief exposures. These exposures were as short as 3 milliseconds. This occurs well before the observer can explicitly say what they saw.
The gap between seeing and knowing
A central challenge in studying rapid perception is separating detection from awareness. In many psychological paradigms, researchers assume that if a participant performs well, they were consciously aware of the stimulus. However, "doing" and "knowing" are not always synchronized.
Previous studies have struggled to decouple these two processes. If a person can locate a face in a dark room, did they "see" it? Or did their brain simply process the signal subconsciously? Without a way to measure the exact moment objective detection occurs versus the moment subjective awareness kicks in, we cannot determine if eye contact is a prioritized signal.
To solve this, the researchers must distinguish between three distinct levels of processing:
Localisation (Detection): Can the participant accurately find where the face appeared on the screen?
Categorisation (Discrimination): Can the participant tell you if the gaze was direct or averted?
Metacognitive Access (Awareness): Does the participant's subjective rating of "how much they saw" actually track their performance?
By varying the exposure from 1 to 5 ms, the authors mapped the arrival of each capability. They also used information-theoretic analyses (mathematical tools used to quantify how much certainty a response provides about a stimulus). This helped them see if physical responses carried information before verbal reports did.
Detecting gaze before the mind wakes up
The results reveal a striking dissociation between sensory processing and conscious experience. The authors find that direct gaze acts as a biological boost for face detection. Specifically, they report that direct-gaze faces require less visual stimulation to be localized than averted-gaze faces. This produces higher sensitivity from 3 ms onwards .
Figure 2: Psychometric evidence for a direct-gaze advantage in face detection at minimal visual stimulation. (A) Exposure duration required to reach 75% correct localisation of the intact face, estimated from psychometric functions fitted separately for each gaze direction and emotional expression condition. Lower thresholds indicate that less visual stimulation was required for detection. Direct-gaze faces showed lower detection thresholds than averted-gaze faces, whereas emotional expression did not significantly affect thresholds. (B) Curve-level localisation accuracy across exposure durations. Accuracy increased as visual stimulation increased from 1 to 5 ms, and was higher for direct gaze than averted-gaze faces, confirming that gaze direction modulated detection performance across minimal exposures. (C) Just-noticeable differences estimated from the fitted psychometric functions. Smaller JNDs indicate greater discrimination precision. Gaze direction significantly affected JNDs, indicating greater precision for direct-gaze than averted-gaze faces. (D) Slope parameter estimates from the psychometric functions. Steeper slopes indicate a sharper increase in localisation performance as exposure duration increased. Slopes were significantly modulated by gaze direction, consistent with enhanced psychophysical performance for direct gaze. Error bars indicate ± 1 SEM. Individual lines in panel B indicate individual participants. Asterisks denote significant differences.
Crucially, this advantage appears in a window where the participant is essentially blind to the social details. The paper finds that while localisation sensitivity (the ability to find the face) is elevated by direct gaze at 3 ms, the ability to categorize the gaze direction does not become reliable until 4 to 5 ms [Figure 4C].
Furthermore, the researchers report that metacognitive sensitivity (the degree to which a person's subjective sense of "seeing" matches their actual accuracy) only emerges at 4 ms or later [Figure 4D]. This creates a clear temporal gap. The brain uses the information from eye contact to orient itself in space (localisation) before the mind realizes it has seen a pair of eyes. The information-theoretic data reinforces this. It shows that localisation responses carry stimulus-location information before subjective awareness ratings become informative .
Figure 5: Information-theoretic evidence that eye contact increases localisation information. (A) Mutual information between the actual location of the intact face and participants' localisation responses, I (stimulus location; localisation response), plotted separately by emotional expression and gaze direction. Localisation responses carried increasing information about stimulus location as exposure duration increased. (B) Direct-gaze advantage in localisation information, computed as the direct-minusaverted difference in localisation mutual information. Direct gaze increased localisation information at 4 ms and 5 ms for both fearful and neutral faces. (C) Co-information between stimulus location, localisation response, and PAS ratings, CoI (stimulus location; localisation response; PAS), shown separately for the four expression-by-gaze conditions, indexing the extent to which subjective awareness ratings overlapped with or modulated localisation information. Reliable positive co-information emerged at 5 ms for fearful-direct, neutral-direct, and neutral-averted faces, indicating that subjective awareness became coupled to localisation information at the longest exposure duration. (D) Mutual information between emotional expression and expression categorisation responses, I (expression shown; expression reported), plotted separately for direct- and averted-gaze faces. (E) Mutual information between gaze direction and gaze categorisation responses, I (gaze direction; gaze direction response), plotted separately for fearful and neutral faces. (F) Specificity contrasts for expression and gaze categorisation information. The expression contrast shows the direct-minus-averted difference in expression information; the gaze contrast shows the fearful-minus-neutral difference in gaze-direction information. Lines indicate group means and shaded areas indicate within-subject SEM. Dotted horizontal lines indicate zero information or zero contrast. Horizontal markers indicate exposure durations at which information estimates were significantly greater than zero after Holm correction. Asterisks indicate significant planned direct-minus-averted contrasts after Holm correction. For PAS co-information, the direct-minus-averted contrast at 5 ms indicates stronger coupling between subjective awareness ratings and localisation information for direct than averted gaze.
Coarse signals and individual differences
The study also investigates the bandwidth of this early signal. In a second experiment, the authors applied spatial-frequency filtering. This technique separates images into coarse shapes and fine details. They wanted to see which type of information supports these ultra-brief windows.
They found that the earliest eye-contact advantage is driven primarily by low-spatial-frequency (LSF) information. This refers to the coarse, blurry outlines of a face. It does not rely on high-spatial-frequency (HSF) details like the fine texture of an iris .
Figure 6: Type-1 sensitivity across spatial frequency conditions in Experiment 2. (A,B) Localisation sensitivity for (A) , low-spatial-frequency, LSF, and (B) , high-spatial-frequency, HSF, faces, indexed by type-1 d ′ for detecting the location of the intact face. Sensitivity increased with exposure duration, was higher for LSF than HSF faces, and was higher for direct-gaze than averted-gaze faces. (C,D) Eye-contact effect, indexed as the direct-minus-averted difference in localisation sensitivity, for (C) , LSF, and (D) , HSF, faces. The direct-gaze advantage was evident in the LSF condition at both exposures, but emerged in the HSF condition only at 5 ms. (E,F) Expression sensitivity for (E) , LSF, and (F) , HSF, faces, indexed by type-1 d ′ for categorising fearful versus neutral facial expressions. Expression sensitivity was reliable only for direct-gaze faces in the LSF condition. (G,H) Gaze direction sensitivity for (G) , LSF, and (H) , HSF, faces, indexed by type-1 d ′ for categorising direct versus averted gaze. Gaze direction sensitivity was comparatively weak, with only limited evidence of above-chance performance. Horizontal lines above the x-axis indicate exposure durations with above-chance sensitivity, or direct-minus-averted differences above zero, after Holm correction. Data are presented as mean ± 1 SEM. All face stimuli, including the intact face shown here, were taken from the Radboud Face Database (RaFD) and are presented as a stimulus example (see: https://rafd.nl/)
This suggests that eye contact is a low-resolution priority. The brain does not need a high-definition image to trigger a social response. It only needs the coarse structural cue that someone is looking. For researchers studying rapid social processing, this implies that focusing on LSF components may be necessary to capture these earliest signals.
However, the authors note several important caveats. First, the association between autistic traits and a reduced eye-contact advantage was observed in a neurotypical sample. It also did not survive certain statistical corrections. Therefore, the link to autism requires more rigorous clinical validation. Second, while the study proves that eye contact facilitates detection, it does not pinpoint the exact neural architecture. It remains unknown if the signal travels through subcortical pathways (fast, automatic routes) or early cortical regions (more complex processing).
The verdict
The evidence strongly supports the conclusion that eye contact is a high-priority, low-latency signal. The authors have demonstrated that the visual system extracts social relevance from remarkably sparse sensory input. This happens entirely outside the realm of conscious awareness for the first few milliseconds of exposure.
For those studying how biological systems handle rapid signal prioritization, this work provides a clear takeaway. Social salience is not a product of conscious thought. Instead, it is a primitive feature of the visual stream. It begins with coarse, low-frequency information that influences behavior before the mind even registers the encounter.
Figures from the paper
Figure 1: Schematic trial procedure of Experiment 1. (A) LCD tachistoscope used to present stimuli at ultra-brief exposure durations. The apparatus contains two LCD screens arranged orthogonally and combined optically through a semipermeable mirror. The horizontal screen presented target stimuli and its backlight was controlled by a dedicated microcontroller, enabling stimulus durations as brief as 0.002 ms. The vertical screen presented fixation displays, placeholders, response cues, and other non-target elements at regular presentation durations. (B) Participants viewed the superimposed display through a viewing aperture positioned 57 cm from the stimulus plane. (C) Face stimuli varied in emotional expression, fearful or neutral, and gaze direction, direct or averted. (D) On each trial, an intact face and its phase-scrambled counterpart were presented for a predefined exposure duration, 1 to 5 ms. Participants reported the location of the intact face, 2AFC, left or right, categorised its emotional expression, fearful or neutral, categorised its gaze direction, direct or averted, and rated their subjective visual experience of the intact face using a perceptual awareness scale, PAS. All face stimuli, including the intact face shown here, were taken from the Radboud Face Database (RaFD) and are presented as a stimulus example (see: https://rafd.nl/)Figure 3: Extending the detection-discrimination dissociation logic to metacognitive access. Behavioural evidence can be used to distinguish conscious from unconscious contributions to detection effects. Stimulus information contains a critical dimension, { x 1 , x 2 } , such as upright versus inverted faces or direct versus averted gaze. This information may influence behaviour through conscious and/or unconscious perceptual processing. (A) In a standard detection paradigm, a positive detection difference between conditions, ∆ Detection, indicates that { x 1 , x 2 } affects detection, but this alone does not reveal whether the effect depends on conscious or unconscious processing. (B) The detection-discrimination dissociation paradigm addresses this ambiguity by adding a direct discrimination measure for the critical dimension. If both ∆ Detection and discrimination are above chance, conscious processing cannot be ruled out as the source of the detection effect. (C) If ∆ Detection is above chance while discrimination of { x 1 , x 2 } remains at chance, the detection effect can be attributed to unconscious processing of the critical dimension. (D) We extend this logic by adding a metacognitive-access measure. In this extended framework, detection indexes whether the critical dimension influences behaviour, discrimination indexes whether that dimension is available for explicit report, and metacognitive access indexes whether subjective ratings track the perceptual evidence supporting detection performance. (E) The strongest evidence for unconscious processing is obtained when ∆ Detection is above chance, but both discrimination and metacognitive access remain at chance. This pattern indicates that the critical stimulus dimension guides detection behaviour while remaining unavailable to explicit discrimination, and while subjective ratings fail to distinguish correct from incorrect detection responses. The right-hand columns show idealised data patterns expected for unconscious and conscious effects. Points and error bars indicate estimated effects and their uncertainty, with values around zero reflecting chance-level performance or absent metacognitive sensitivity. Greyed elements indicate pathways or measures that are unavailable or non-informative in the corresponding idealised pattern. This figure extends schematics previously presented by Schmidt and Vorberg[38], and Stein and Peelen[37]. 5Figure 4: Type-1 and type-2 sensitivity across minimal exposure durations in Experiment 1. (A) Localisation sensitivity, indexed by type-1 d ′ , for detecting the position of the intact face. Sensitivity increased with exposure duration and was higher for direct gaze than averted-gaze faces from 3 ms onwards. (B) Expression sensitivity, indexed by type-1 d ′ , for categorising fearful versus neutral facial expressions. Expression sensitivity increased with exposure duration and became reliably above chance at longer exposures. (C) Gaze direction sensitivity, indexed by type-1 d ′ , for categorising direct versus averted gaze. Gaze categorisation improved with increasing visual stimulation and became reliable only at longer exposure durations. (D) Metacognitive sensitivity, estimated using maximum likelihood metad ′ , based on participants' perceptual awareness ratings. This measure quantifies the extent to which subjective experience tracked objective perceptual evidence across exposure durations. Metacognitive sensitivity increased with exposure duration and was higher for direct gaze than averted-gaze faces from 4 ms onwards. (E) Metacognitive efficiency, estimated using Bayesian hierarchical M-ratio. Higher values indicate a closer correspondence between type-1 sensitivity and metacognitive access to that evidence. Posterior contrasts were assessed using 95% highest-density intervals, HDI. Contrasts whose confidence intervals included zero suggest no reliable difference between conditions. Metacognitive access to type-1 evidence was higher for direct-gaze faces from 4 ms onwards. Horizontal lines on top of the x-axis of Panels A-D indicate exposure durations with above-chance sensitivity, p < 0 . 05 , one-sample t -tests against zero, Holm-corrected. Data are presented as mean ± 1 SEM, with shaded areas indicating SEM. Asterisks indicate statistically significant effects or planned comparisons after correction for multiple comparisons.