Lead
Researchers have found that some deep neural networks (DNNs) fall for the same optical illusions that trick human observers, a discovery reported in December 2025. Experiments using models such as PredNet — trained on roughly one million video frames of natural scenes — show similar misperceptions for motion and perspective illusions like the rotating snakes. The parallels give new, ethically permissive ways to test hypotheses about visual processing, including predictive coding, while also highlighting clear differences such as attention and spatial focus. Together, the findings are reshaping how neuroscientists link algorithmic models to human perception.
Key takeaways
- PredNet, a predictive deep neural network trained on ~1 million natural frames, was susceptible to the rotating snakes illusion, mirroring human reports of perceived motion in a static image.
- Not all illusions transfer: no single DNN yet reproduces the full range of human illusion experiences, indicating gaps in current architectures.
- Case studies of humans with altered visual development show shape-based illusions can be more impacted by sensory history than motion-based illusions, suggesting different developmental windows.
- Researchers used AI to probe predictive coding theories: the model’s errors resemble expectations-driven misperceptions, supporting predictive accounts of vision.
- PredNet lacks an attention mechanism and therefore treats all image regions uniformly; humans instead show focal suppression (e.g., a fixated rotating disc appears to stop).
- Separately, a quantum-inspired AI model reproduced bistable switches (Necker cube, Rubin vase) with timing similar to humans, without implying brain quantum states.
- Astronaut studies show gravitational context alters illusion balance after ~3 months in orbit, raising operational concerns for long-duration spaceflight.
Background
Optical illusions reveal that visual perception is not a literal recording of the world but an interpretive process that emphasises salient structure over exhaustive detail. Classic examples include the Moon appearing larger on the horizon, the Kanizsa square that creates illusory contours, and motion illusions like the rotating snakes. Scientists study such effects to infer the heuristics and shortcuts the brain uses to interpret ambiguous input.
Deep neural networks (DNNs) emulate aspects of hierarchical visual processing and have excelled at tasks such as medical-image screening by detecting subtle patterns. Because artificial systems can be probed freely and manipulated without ethical constraints, researchers view them as valuable experimental tools to test models of human perception. Still, DNNs are engineered with objectives and architectures that differ from biological nervous systems, so parallels must be interpreted cautiously.
Main event
In a recent set of experiments, Eiji Watanabe and colleagues trained PredNet, a predictive coding–inspired network, on videos taken from head-mounted cameras to approximate the statistics of human visual experience. After learning temporal regularities from roughly one million frames, the model was shown static frames of the rotating snakes illusion and some human-resistant variants. PredNet produced internal prediction errors consistent with a perception of motion for the same images that fool human observers.
The researchers interpret these errors as the model internally signaling a mismatch between expectation and input — a process analogous to predictive coding theories where the brain generates expectations and then processes residual errors. Crucially, PredNet had never been exposed to illusions during training; the misperception emerged from its learned rules about motion and appearance in natural scenes.
Differences between model and human behaviour also emerged. Humans typically report that a disc under direct fixation in the rotating snakes pattern seems to stop while peripheral discs continue to spin; PredNet, which lacks an attention module, indicated motion across the entire image simultaneously. This divergence highlights attention and eye-movement dynamics as likely missing components in many DNNs.
Separately, Ivan Maksymov at Charles Sturt University explored a quantum-inspired approach to bistable illusions such as the Necker cube and Rubin vase. His model produced spontaneous switching between the two perceptual interpretations at time intervals resembling those measured in people, suggesting that certain stochastic or non-classical frameworks can capture aspects of perceptual ambiguity without asserting quantum biology.
Analysis & implications
The convergence of AI misperception and human illusion offers mutual benefits: models provide testable implementations of theoretical accounts (for example, predictive coding), and human data highlight which computational features remain missing in models. The fact that PredNet displayed the rotating snakes effect after passive exposure to natural scenes supports the idea that expectations about motion statistics can produce illusory motion when local cues mimic moving objects.
However, similarity in behaviour does not prove identical mechanisms. PredNet’s errors arise from objective optimization of frame prediction; human perception is shaped by attention, embodied constraints, and developmental history. The absence of gaze-driven modulation in PredNet suggests that adding attention, saccade models or recurrent constraints may be necessary to bridge the remaining gap.
Practical implications span neuroscience, AI safety and spaceflight. For neuroscience, AI models give a manipulable substrate to falsify mechanistic claims that would be ethically or technically infeasible in humans. For AI safety, understanding when and why systems misperceive can inform robustness testing in safety-critical domains like autonomous vehicles and medical imaging. For space exploration, shifting illusion dynamics under microgravity — where depth cues tied to gravity change — raises questions about perceptual reliability for crew during long missions.
Comparison & data
| Illusion | Human response | AI (example) | Notes |
|---|---|---|---|
| Rotating snakes | Perceived motion in static pattern | PredNet predicts motion (shows prediction error) | PredNet trained on ~1,000,000 frames of natural video; fooled by the same variants as humans |
| Necker cube / Rubin vase | Bistable switching between interpretations | Quantum-inspired model showed switching with human-like timing | Model used quantum tunnelling-inspired computation; not evidence of brain quantum states |
| Kanizsa square (illusory contour) | Perceived contours where none exist | Some DNNs can encode contour-like features, but responses vary by architecture | Human susceptibility depends on developmental exposure; DNN training data matters |
| Barber pole (motion) | Perceived vertical motion for rotating diagonal stripes | Motion-trained models often reproduce direction ambiguity | Motion perception appears more resilient to sensory deprivation than shape perception |
The table summarises qualitative parallels observed in recent studies. While numerical metrics differ across experiments, a consistent pattern emerges: models that learn temporal regularities can reproduce motion illusions; bistability requires additional stochastic or recurrent dynamics; and attention/gaze effects are commonly missing from standard architectures.
Reactions & quotes
Using DNNs allows controlled tests of how learned expectations shape perception, without the ethical limits of human brain manipulation.
Eiji Watanabe, National Institute for Basic Biology (paraphrased)
Watanabe emphasises that artificial models make it possible to run manipulations that would be impractical or unethical in people, enabling direct checks of mechanistic hypotheses such as predictive coding.
We can model bistable perception with quantum-inspired computations, but this does not imply the brain literally operates on quantum mechanics.
Ivan Maksymov, Charles Sturt University (paraphrased)
Maksymov notes that quantum-inspired architectures can capture the switching dynamics of ambiguous figures while remaining agnostic about underlying biology.
Unconfirmed
- Whether predictive coding fully explains all classes of optical illusions remains unproven; current evidence supports the theory for some motion illusions but not all shape or bistable phenomena.
- Connections between quantum-inspired AI models and biological quantum processes are speculative; similar behaviour in an algorithm does not confirm quantum mechanisms in brains.
- The exact role of attention and eye movements in producing differences between humans and DNNs is inferred from behavioural divergences but not definitively mapped in these specific models.
Bottom line
AI models are proving to be powerful, manipulable tools for probing visual perception. When networks trained on naturalistic visual input reproduce human-like illusion effects, they lend support to algorithmic accounts — such as predictive coding — that emphasise expectation-driven perception. At the same time, systematic differences remind us that current DNNs omit critical features of biological vision: attention, active sensing (eye movements) and developmental constraints.
For scientists, engineers and space agencies, these findings are both a resource and a caution. Models can accelerate hypothesis testing and expose vulnerabilities in deployed systems, but overinterpreting superficial similarity risks missing the full complexity of embodied human perception. The next steps are to integrate attention, oculomotor behaviour and richer developmental priors into models, and to test predictions in both lab and operational settings including spaceflight.