Is AI-Generated Content Silently Rewiring Our Visual Brain?

By Dr. Hao Liu, Assistant Professor, School of Computing, Montclair State University

The Question That Keeps Me Up at Night

Here’s a thought experiment. You spend two hours scrolling through a feed filled with AI-generated images — photorealistic faces, landscapes, architectural renders, short videos. You can’t tell which ones are real. Neither can most people: recent research shows human accuracy at detecting AI-generated images has dropped to as low as 29% for state-of-the-art models (Exner et al., 2025).

Now here’s the uncomfortable question: what is your visual brain doing with all of that?

The Latent Space Mismatch Hypothesis

Our visual system is, in computational terms, a generative model trained over millions of years of evolution on a single dataset: physical reality. It has learned the statistical regularities of how light behaves, how objects occupy space, how biological motion unfolds, how causes lead to effects. These learned priors form what we might call the brain’s internal latent space — a compressed representation of the world that enables everything from catching a ball to reading facial expressions.

AI generative models — diffusion models, GANs, autoregressive video generators — learn a different distribution. Their latent space is trained on internet-scale data through vector-based sampling. The output approximates reality, but it is not reality. The deviations are subtle: physically inconsistent shadows, texturally uncanny skin, temporally implausible motion sequences, causally disconnected event chains.

Here is my core hypothesis: prolonged, large-scale exposure to photorealistic AI-generated content may cause a non-adaptive drift in the human visual system’s internal priors — a gradual, subthreshold modification of perceptual representations that were calibrated to the statistics of the physical world.

Why This Isn’t Just Another “Screen Time Is Bad” Argument

I want to be precise about what makes this different from previous concerns about media consumption:

Scale. We’re not talking about occasional movie-watching. We’re talking about hours per day of AI-generated content mixed seamlessly into information feeds. The generative AI market is projected to grow 560% between 2025 and 2031 (Elsner et al., 2025), and we are approaching what UNESCO has termed a “synthetic reality threshold” — a point beyond which humans can no longer reliably distinguish authentic from fabricated media without technological assistance (UNESCO, 2025).

Fidelity. An impressionist painting doesn’t fool your low-level visual system. A photorealistic AI-generated scene does. The deviations operate below the threshold of conscious detection but within the operating range of unconscious statistical learning mechanisms. Research has shown that humans detect AI-generated faces at roughly chance level, and even for video deepfakes, performance on high-quality samples can drop below chance (Korshunov & Marcel, 2021; Lago et al., 2022).

Unlabeled mixing. When you watch a Marvel movie, your brain’s source monitoring system tags it as fiction. When AI content is mixed unlabeled into your social media feed, this protective mechanism is bypassed. Studies show that sustained exposure to AI-mediated environments gradually shifts decision-making heuristics away from the detection of discrete artifacts and toward reliance on vague “uncanny” intuitions (Di Plinio, 2025), suggesting the perceptual system is already adapting — but not necessarily in a beneficial direction.

These three factors converging simultaneously is historically unprecedented.

What the Existing Literature Tells Us

This hypothesis sits at an intersection where several established research streams converge, but no one has connected the dots:

Unconscious statistical learning. The human brain extracts statistical regularities from sensory input automatically and without conscious awareness. This was first demonstrated in the auditory domain by Saffran, Aslin, and Newport (1996), who showed that 8-month-old infants could learn transitional probabilities in speech streams after just two minutes of passive listening. Fiser and Aslin (2001, 2002) extended these findings to vision, showing that adults and infants spontaneously learn the statistical structure of visual scenes — spatial co-occurrence patterns, conditional probabilities between visual elements — without instruction or task engagement. Critically, this learning is automatic: it occurs even for behaviorally irrelevant stimuli that are not consciously attended to (Turk-Browne et al., 2005). This means the visual system will encode whatever statistical regularities it is exposed to, whether they come from the physical world or from an AI generative model.

Visual adaptation neuroscience. Long-term exposure to non-standard visual input produces lasting changes in perceptual processing. Webster (2015) provides a comprehensive review showing that visual adaptation operates across timescales from milliseconds to years, selectively recalibrating sensitivity. Long-term exposure to colored biases induces persistent aftereffects far exceeding those from short-term adaptation (Neitz et al., 2002; Belmore & Shevell, 2010), and color vision in cataract patients requires months to readapt after surgery (Delahunt et al., 2004). Most striking is the work on keratoconus patients: Vienola et al. (2021) showed that chronic exposure to severe optical aberrations causes a functional reallocation of sensory processing resources — impaired fine spatial vision alongside enhanced sensitivity to coarse spatial information — demonstrating that the adult visual system genuinely recalibrates its contrast sensitivity function in response to prolonged non-standard input.

Passive exposure learning. Beyeler et al. (2022) demonstrated that repetitive passive exposure to oriented stimuli can induce a persistent, bottom-up form of perceptual learning that is stronger than top-down practice-based learning, with broader generalization to complex stimuli including natural scenes. This is critical: it means the visual system doesn’t need to be “trying” to learn from AI-generated content. Mere exposure is sufficient to modify perceptual representations.

AI content and false memory. A landmark CHI 2025 study by Pataranutaporn et al. (2025), “Synthetic Human Memories,” demonstrated that AI-edited images and videos can implant false memories and distort recollection of past events. In their experiment with 200 participants, AI-modified visuals significantly altered participants’ memories of original scenarios, with edits involving people producing the most pronounced effects. If AI content can modify explicit episodic memory, the question of whether it modifies implicit perceptual priors is a natural and urgent next step.

Deepfake cognitive impact. A 2025 scoping review by Greiner et al. in AI & Society systematically examined 28 empirical studies on the harms of deepfakes, identifying multiple categories: false memories, attitude shifts, sharing intention changes, media distrust, anxiety, reduced self-efficacy, and distress from sexual deepfake victimization. A separate scoping review by Murphy et al. (2025) in PLOS ONE focused specifically on deepfakes’ effects on beliefs, memories, and behavior, noting that it remains unknown whether deepfakes are more, less, or equally effective compared to text- or photo-based misinformation. Crucially, nearly all existing research operates at the information level (misinformation, trust, memory for events). The perceptual level — whether low-level visual priors are modified — remains unexplored.

The “Impostor Bias” phenomenon. The proliferation of deepfakes has produced a novel cognitive bias: individuals increasingly question the authenticity of all multimedia content, even when it is genuine (Amerini et al., 2025). This suggests that mere awareness of AI-generated content is already modifying how people process visual information, even before we consider the subthreshold effects of exposure to the content itself.

Predictive coding and the free energy principle. Friston and Kiebel (2009) formalized perception as hierarchical Bayesian inference, where the brain continuously generates top-down predictions and updates its internal models based on bottom-up prediction errors. The precision (inverse variance) assigned to prediction errors determines how strongly they update the model. Exposure to content that unpredictably violates physical priors — sometimes consistent with reality, sometimes subtly not — could disrupt this precision weighting mechanism. The brain would face a non-stationary statistical environment where its confidence calibration becomes unreliable. Disrupted precision weighting is, notably, a proposed mechanism in clinical conditions including psychosis and derealization (Clark, 2013; Hohwy, 2013).

Images vs. reality in the brain. Culham and colleagues (2021) reviewed a growing body of evidence that both behavior and brain function differ between image proxies and real, tangible objects, with differences found in perception, memory, and attention. They note that “evolution and development are shaped by the real world” and that shortfalls between proxies and reality are particularly evident in other species and young children. This establishes the broader principle that the visual system is calibrated to physical reality, not to images — making it vulnerable when images become indistinguishable from reality.

The Gap — and the Opportunity

To my knowledge, no published study has directly measured whether long-term exposure to AI-generated photorealistic content alters low-level visual priors — sensitivity to lighting consistency, biological motion plausibility, texture statistics, or causal coherence.

This is the gap I want to fill.

The Research Plan

I’m designing a multi-phase research program:

Phase 1 — Psychophysics Pilot. Quantify the low-level statistical differences between AI-generated and real images (power spectrum, illumination consistency via spherical harmonics, texture statistics following Portilla & Simoncelli, 2000). Test whether these differences produce measurable subthreshold perceptual effects using oddity detection and adaptation aftereffect paradigms.

Phase 2 — Longitudinal Exposure Study. Expose participants to controlled daily doses of AI-generated vs. real visual content over 4–8 weeks. Measure weekly changes in physical consistency sensitivity, biological motion judgment, and causal sequence evaluation. Optionally incorporate EEG (visual mismatch negativity as a neural index of predictive coding; Stefanics et al., 2014) or fMRI adaptation paradigms to track representational changes at the neural level.

Phase 3 — Computational Modeling. Build a hierarchical predictive coding model following the free-energy formulation (Friston & Kiebel, 2009; Bogacz, 2017), train it on natural images, then expose it to AI-generated content and track representational drift. Compare model predictions with behavioral and neural data from Phase 2.

Who I’m Looking For

I’m actively seeking collaborators with expertise in:

• Visual psychophysics / visual neuroscience — experimental design for threshold measurement, adaptation paradigms, stimulus control
• Predictive coding / free energy principle — computational modeling of perceptual inference under non-stationary statistics
• Neuroimaging (fMRI/EEG) — particularly experience with visual mismatch negativity, repetition suppression, or adaptation paradigms
• Computational cognitive science — modeling perceptual learning and representational drift

I bring expertise in AI/ML systems, deep learning architectures, and medical informatics, along with institutional resources and a track record in interdisciplinary research. What I need are partners who understand the perceptual neuroscience side deeply enough to ensure experimental rigor.

Why This Matters Beyond Academia

If prolonged AI content exposure does shift human perceptual priors — even modestly — the implications extend far beyond the lab:

• Clinical: Could this contribute to rising rates of derealization and perceptual instability, particularly in heavy digital media consumers?
• Developmental: What happens when children’s visual systems are calibrated not on physical reality but on a mixture of real and AI-generated statistics?
• Societal: If causal reasoning thresholds drop, does susceptibility to false narratives and conspiracy thinking increase?
• Regulatory: Should AI-generated content carry mandatory labeling not just for misinformation reasons, but for perceptual health?

These are empirical questions. Let’s answer them.

Get in Touch

If this resonates with your research interests, I’d love to hear from you. Whether you’re a visual neuroscientist intrigued by the perceptual implications, a computational modeler interested in predictive coding under distributional shift, or a cognitive psychologist studying media effects — there’s a meaningful contribution to be made here.

References

Amerini, I., Barni, M., Battiato, S., Boato, G., Ferrara, P., Guilaro, E., … & Verdoliva, L. (2025). Deepfake media forensics: State of the art and challenges ahead. In Advances in Social Networks Analysis and Mining (ASONAM 2024), Lecture Notes in Social Networks. Springer.

Belmore, S. C., & Shevell, S. K. (2010). Very-long-term chromatic adaptation: Test of gain theory and a new method. Visual Neuroscience, 28(1), 17–24.

Beyeler, M., Bhatt, U., & Bhalla, U. S. (2022). Learning by exposure in the visual system. Current Biology, 32(9), R431–R436.

Bogacz, R. (2017). A tutorial on the free-energy framework for modelling perception and learning. Journal of Mathematical Psychology, 76, 198–211.

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.

Culham, J. C., Cavina-Pratesi, C., & Engel, S. A. (2021). The treachery of images: How realism influences brain and behavior. Trends in Cognitive Sciences, 25(6), 506–519.

Delahunt, P. B., Webster, M. A., Ma, L., & Werner, J. S. (2004). Long-term renormalization of chromatic mechanisms following cataract surgery. Visual Neuroscience, 21(3), 301–307.

Di Plinio, S. (2025). Perceived agency and AI-mediated environments. [Cited in Exner et al., 2025].

Elsner, C., et al. (2025). Global Risks Report 2025. World Economic Forum.

Exner, Y., Hartmann, J., & Domdey, S. (2025). What you see is not what you get anymore: A mixed-methods approach on human perception of AI-generated images. Frontiers in Artificial Intelligence, 8, 1707336.

Fiser, J., & Aslin, R. N. (2001). Unsupervised statistical learning of higher-order spatial structures from visual scenes. Psychological Science, 12(6), 499–504.

Fiser, J., & Aslin, R. N. (2002). Statistical learning of new visual feature combinations by infants. Proceedings of the National Academy of Sciences, 99(24), 15822–15826.

Friston, K., & Kiebel, S. (2009). Predictive coding under the free-energy principle. Philosophical Transactions of the Royal Society B, 364(1521), 1211–1221.

Greiner, A., et al. (2025). The harm of deepfakes: A scoping review of deepfakes’ negative effects on human mind and behavior. AI & Society. https://doi.org/10.1007/s00146-025-02774-0

Hohwy, J. (2013). The Predictive Mind. Oxford University Press.

Korshunov, P., & Marcel, S. (2021). Deepfake detection: Humans vs. machines. arXiv preprint arXiv:2009.03155.

Lago, F., Pasquini, C., Böhme, R., Dumont, H., Goffaux, V., & Boato, G. (2022). More real than real: A study on human visual perception of synthetic faces. IEEE Signal Processing Magazine, 39(1), 109–116.

Murphy, G., et al. (2025). Can deepfakes manipulate us? Assessing the evidence via a critical scoping review. PLOS ONE, 20(5), e0320124.

Neitz, J., Carroll, J., Yamauchi, Y., Neitz, M., & Williams, D. R. (2002). Color perception is mediated by a plastic neural mechanism that is adjustable in adults. Neuron, 35(4), 783–792.

Pataranutaporn, P., Leong, J., Danry, V., Lawson, A. P., Maes, P., & Sra, M. (2025). Synthetic human memories: AI-edited images and videos can implant false memories and distort recollection. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM.

Portilla, J., & Simoncelli, E. P. (2000). A parametric texture model based on joint statistics of complex wavelet coefficients. International Journal of Computer Vision, 40(1), 49–71.

Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926–1928.

Stefanics, G., Kremláček, J., & Czigler, I. (2014). Visual mismatch negativity: A predictive coding view. Frontiers in Human Neuroscience, 8, 666.

Turk-Browne, N. B., Jungé, J. A., & Scholl, B. J. (2005). The automaticity of visual statistical learning. Journal of Experimental Psychology: General, 134(4), 552–564.

UNESCO. (2025). Deepfakes and the crisis of knowing. https://www.unesco.org/en/articles/deepfakes-and-crisis-knowing

Vienola, K. V., et al. (2021). Functional reallocation of sensory processing resources caused by long-term neural adaptation to altered optics. eLife, 10, e58734.

Webster, M. A. (2015). Visual adaptation. Annual Review of Vision Science, 1, 547–567.

Hao is an Assistant Professor in the School of Computing at Montclair State University. His research focuses on AI applications in healthcare and the cognitive implications of generative AI systems.