Illusions can fool image-recognition networks too. In an early and influential study, a Japanese neurophysiology team fed deep neural networks a stationary image of the famous "rotating snakes" illusion — the one where static circles appear to spin.The networks, despite having no eyes, no eye movements, and no biology at all, "perceived" motion in the same static image, in ways that lined up with predictive coding theory — the idea that a visual system isn't passively recording pixels but constantly predicting what comes next and getting fooled when the prediction overrides the raw signal.
.png)
That prediction-engine framing keeps showing up. One recent research summary put it well: certain AI models misjudge size, brightness, and motion in classic optical illusions, and in some cases the AI's mistake mirrors human error almost perfectly — for instance, contrast illusions cause AI to perceive false boundaries or exaggerated differences, just as they do in humans. The takeaway researchers draw from this is that perception errors aren't random — they emerge from the logic of interpretation itself, and when a system tries to make sense of limited data quickly, it will be wrong in consistent, explainable ways. In other words: the fact that AI can replicate human perceptual errors suggests our brains work less like cameras and more like prediction engines that constantly guess what they're seeing and update as new information arrives, with illusions being the cases where those predictions overpower raw sensory input.
A separate line of work went further and tried to build a network that could handle ambiguous illusions — the Necker cube (a wireframe cube that flips between two orientations) and Rubin's vase (the face/vase image). Researcher Ivan Maksymov built what he called a "quantum-tunneling" neural network specifically because ordinary computer vision systems tend to lock onto one interpretation and stay there, whereas humans flicker back and forth between the two readings. His network, borrowing math from quantum tunneling, reproduced that flickering — simulating human perception of the Necker cube and Rubin's vase illusions better than some much larger conventional computer-vision networks. His argument is essentially that when a human sees an image with two valid interpretations, the brain briefly holds both at once before settling on one, and that this superposition-like state is what standard feedforward networks lack and what made them bad at ambiguous illusions.
Not everything lines up neatly, though, and the mismatches are arguably just as informative as the matches. Testing several vision-language models against a large illusion benchmark, one team found the AIs were not particularly good at identifying illusory elements in an image, averaging under 36% accuracy, and their answers matched human responses in only about 16% of cases — though larger models aligned more closely with human perception than smaller ones did. And more broadly, researchers have noticed that not all AI systems fall for the same illusions humans do — some are immune to tricks that reliably fool people, while being vulnerable to others humans barely notice.
That divergence is the more useful half of the story. If perception were just "whatever pattern-matching a sufficiently trained system converges on," you'd expect illusions to hit humans and AI more or less identically. They don't. The differences show that perception depends heavily on architecture and training, and when AI diverges from humans it helps isolate which parts of human vision are learned through lived experience versus which may be built in by evolution — hardwired shortcuts baked in by millions of years of needing to detect motion, edges, and depth cheaply and fast, rather than learned purely from exposure to images the way a neural network is.
There's also a nice bit of context for why the human side of this works the way it does. Neuroscientist Susana MartÃnez-Conde's research on the rotating snakes illusion found it isn't just about the image itself — the illusion requires a triggering event, specifically involuntary eye movements called saccades or microsaccades, and if you don't move your eyes, the illusion doesn't work. It's also not universal: younger people are considerably more likely than older people to perceive the motion at all. That's a reminder that even the "biological baseline" AI is being compared against isn't a fixed, single thing — it varies with eye physiology and age, which makes the AI comparison messier and more interesting than a simple "does the machine see it too" checkbox.
Put together, the practical upshot researchers keep coming back to is that illusions are turning into a diagnostic tool in both directions — for probing what's "hardwired" versus learned in human vision, and for stress-testing AI vision systems in domains like autonomous driving or medical imaging where a confident, illusion-style misread of an image would have real consequences.