Separating what a face is from what it is called
Most earlier work could not tell two things apart: the effect of an image actually being synthetic, and the effect of a viewer being told it is synthetic. Studies that only swap the image leave belief to chance; studies that only swap a label keep the image the same and therefore cannot test a label that contradicts the face. This paper crosses both factors inside one experiment, so a response can be attributed to the face, to the label, or to their agreement.
The design has two stages. An online study (N=169, 6,760 valid trials) had raters score diffusion-generated faces for perceived AI-likeness on a 0–100 scale; only 82 of 320 candidate faces whose lower 95% confidence bound exceeded 60 were kept, so the lab stimuli would be reliably read as synthetic rather than ambiguous. The lab study then showed faces from two actual sources (AI-generated, human) under three labels (no label, "Made with AI", "Made by a human"), recording EEG and gaze from 30 participants.
Crossing the factors produces congruent pairings (an AI face labelled "Made with AI") and incongruent ones (an AI face labelled "Made by a human"). That incongruent cell is the part earlier designs could not reach, and it is where the most distinctive result appears.
Three ERP windows track attention, re-evaluation, and affect
Event-related potentials (ERPs) are averaged voltage changes time-locked to the image, and each named component indexes a different processing stage. The N2, measured over frontocentral electrodes in a 200–350 ms window, is treated here as an early attention response. The P3, over centroparietal electrodes in a 300–500 ms window, is linked to re-evaluating a stimulus. The LPP, in a 400–800 ms window, indexes sustained affective processing.
Because the paper's central claim is an absence of difference between AI and human faces, the analysis uses a Bayesian hierarchical model rather than a threshold test. A posterior concentrated near zero states an absence directly, whereas a non-significant test cannot distinguish no effect from no evidence. All analyses were run at the trial level, with participant-specific slopes for the experimental factors and by-image intercepts, so estimates are not tied to the particular faces and people observed.
One coding detail matters for reading the numbers. ERP models compare congruent and incongruent labels each against unlabelled faces, while eye-tracking models code label as present versus absent. The two modalities therefore answer slightly different questions and their effect sizes are not interchangeable.
The N2 responds to a label being present
Actual source did not modulate the N2: AI- and human-sourced faces produced equivalent frontocentral amplitudes (median difference 0.01 µV, 95% credible interval [−0.08, 0.11], probability of direction 61.5%). The early response moved with the label instead. Relative to unlabelled faces, labelled faces tended to elicit a larger N2, most clearly when label and face agreed (0.19 µV, [−0.03, 0.40], 95.7%).
The authors read this as sensitivity to the presence of a source cue rather than to conflict between cue and face. The waveform comparison supports that reading: any face carrying a label, matching or not, produced a larger response than an unlabelled face, though the difference is small. Note that the congruent contrast's interval still crosses zero, so this is a directional tendency rather than a firmly bounded effect.
Conflict between label and face registers at the P3
The P3 also showed no difference by actual source (−0.04 µV, [−0.12, 0.04], 81.7%). Here the mismatch did register: incongruent trials, an AI face tagged "Made by a human" or the reverse, produced a larger P3 than unlabelled faces (0.17 µV, [−0.02, 0.36], 95.6%).
This is the paper's sharpest separation. The later evaluative response follows what the label claims relative to what the face is, not the face's true origin. A viewer told the wrong thing about a face engages a re-evaluation stage that a correctly labelled or unlabelled face does not trigger to the same degree. As with the N2, the interval's lower bound sits just below zero, so the effect is credible in direction but not tightly estimated.
Sustained affective processing does not follow the label
The LPP was unmoved by either factor. Neither actual source (−0.01 µV, [−0.09, 0.07], 60.0%) nor congruence (all absolute median differences ≤0.08 µV) modulated it, and the six condition traces overlap across the window. Whatever the label did to earlier processing did not carry into this later affective stage.
This matters for how the result should be described. The study supports a claim about attention and re-evaluation, not about how much a viewer likes or trusts a face. Earlier work cited in the paper had linked believing a face is fake to dampened affective responses; here, with actual and attributed origin separated, that dampening does not appear in the LPP.
Labels change exploration but not when gaze first lands
First-fixation latency showed no credible effect of source (0.316, [−14.593, 15.573], 51.8%), label (−7.397, [−26.664, 11.384], 78.6%), or their interaction. Initial orienting to the face was therefore unchanged, consistent with the idea that labels act on later processing rather than on where the eyes go first.
Exploration itself did shift. Labelled faces were fixated more often than unlabelled ones (−0.043, [−0.064, −0.024], 100%), and those extra fixations were shorter: mean fixation duration fell with a label present (0.053, [0.038, 0.068], 100%) and was also shorter on AI faces than human faces (0.022, [0.010, 0.035], 100%). The increase in fixation count was larger for AI-generated faces than for human faces (0.041, [0.013, 0.070], 99.8%).
In plain terms, participants looked at labelled faces more times without spending longer on them overall. Total dwell time is the product of fixation count and mean duration, and the paper states the three measures are effectively collinear in these data, so dwell time restates the fixation-count interaction rather than adding independent evidence.
Where the stimulus design limits the conclusion
The two stimulus sets were not matched on low-level visual properties. Human faces came from the Chicago Face Database with white backgrounds and gray clothing; AI faces came from a Stable Diffusion collection with their own regularities, such as pastel backgrounds and smoothed skin. The authors name this as an alternative explanation for the shorter fixations on AI faces and the larger label effect for them, so those two gaze results should be read as confounded with appearance. Perceived AI-likeness can also vary with apparent race, which was not controlled.
Study 1 fixed perceived origin by keeping only faces reliably rated as AI-generated, which removed ambiguity but also meant perceived and actual origin were not fully crossed. Whether labels still steer responses when an AI face looks convincingly human remains untested here. The sample was 30 participants in the lab, relatively young with moderate self-reported AI literacy, and the stimuli were static faces under free viewing with one label design in one position, unlike feeds where labels compete with surrounding content.
Why the label, not the pixels, is the lever
The practical priority this study supports is narrow and specific. Because actual source produced no credible effect in any ERP component or in first-fixation latency, while label presence and label–face conflict did move the N2 and P3, the informative variable for disclosure design is the attributed origin rather than the image's true origin. A label is not merely a caption a viewer can weigh and discard; it reaches processing stages the viewer cannot introspect.
The paper's own framing sharpens the stakes: if a label only changed what viewers report, they could inspect it and set it aside; if it changes how the face is processed, a wrong label produces an impression the viewer has no way to catch. The incongruent condition is the direct evidence for the second case, and it is the condition a disclosure regime should most worry about, since a mislabelled face is exactly where re-evaluation is triggered.
Two boundaries keep this from being a general claim about disclosure effectiveness. The affective window stayed flat, so nothing here shows labels changing how faces are felt about, only how they are attended to and re-evaluated. And the gaze differences remain entangled with unmatched stimulus appearance. The next test the authors point to is crossing perceived origin, actual origin, and label, including AI faces that look convincingly human.
Sources used in this answer
EEG and Eye-Tracking Evidence That AI Disclosure Shapes Face Evaluation
Actual source (AI vs human) did not modulate N2, P3, LPP, or first-fixation latency, but label presence enlarged the N2 and label–face conflict enlarged the P3, while the LPP stayed flat and labels increased fixation counts with shorter fixations.