Artificial intelligence

Our eyes often play tricks on us, but scientists have discovered that some artificial intelligence (AI) systems can fall for the same illusions and this is reshaping how we understand the human brain.

Take the Moon, for example. When it’s near the horizon, it appears larger than when it’s high in the sky, even though its actual size and the distance from Earth remain nearly constant. Optical illusions like this show that our perception doesn’t always match reality. While they are often seen as errors, illusions also reveal the clever shortcuts our brains use to focus on the most important aspects of our surroundings.

In reality, our brains only take in a “sip” of the visual world. Processing every detail would be overwhelming, so instead we focus on what’s most relevant. But what happens when a machine a synthetic mind powered by artificial intelligence encounters an optical illusion?

AI systems are designed to notice details humans often miss. This precision is why they can detect early signs of disease in medical scans. Yet, some deep neural networks (DNNs)the backbone of modern AI are surprisingly susceptible to the same visual tricks that fool us. This opens a new window into understanding how our own brains work.

“Using DNNs in illusion research allows us to simulate and analyze how the brain processes information and generates illusions,” says Eiji Watanabe, associate professor of neurophysiology at Japan’s National Institute for Basic Biology. Unlike human experiments, testing illusions on AI carries no ethical concerns.

No DNN, however, can experience all the illusions humans do. Although theories abound, the reasons we perceive certain illusions remain largely unexplained.

Studying people who don’t perceive illusions provides clues. For instance, one person who regained sight in his 40s after childhood blindness was not fooled by shape illusions like the Kanizsa square, where four circular fragments create the illusion of a square. Yet he could perceive motion illusions, such as the barber pole, where stripes seem to move upward on a rotating cylinder.

These observations suggest that our ability to detect motion is more robust than our perception of shapes perhaps because we process motion earlier in infancy, or because shape recognition is more influenced by experience.

Brain imaging, such as fMRI, has also shown which regions of the brain activate when we see illusions and how they interact. Still, perception is subjective. A famous example is the “dress” photo from 2015, which viewers argued over as blue-and-black or white-and-gold. Such differences make illusions difficult to study objectively.

Now AI offers a new approach. Many AI systems, including chatbots like ChatGPT, use DNNs composed of artificial neurons inspired by the human brain. Watanabe and his colleagues investigated whether a DNN could replicate how humans perceive motion illusions, such as the “rotating snakes” illusion a static pattern of colorful circles that appear to spin.

They used a DNN called PredNet, designed around the predictive coding theory. This theory suggests that the brain doesn’t simply process visual input passively. Instead, it predicts what it expects to see, then compares this to incoming sensory data, allowing faster perception. PredNet works similarly, predicting future video frames based on prior observations.

Trained on natural landscape videos, PredNet had never seen an optical illusion before. After processing about a million frames, it learned essential rules of visual perception including characteristics of moving objects. When shown the rotating snakes illusion, the AI was fooled just like humans, supporting the predictive coding theory.

Yet differences remain. Humans experience motion differently in their central and peripheral vision, but PredNet perceives all circles as moving simultaneously. This is likely because PredNet lacks attention mechanisms it cannot focus on a specific area like the human eye.

Even though AI can mimic some aspects of vision, no DNN fully experiences the range of human illusions. “ChatGPT may converse like a human, but its DNN works very differently from the brain,” Watanabe notes. Some researchers are even exploring quantum mechanics to better simulate human perception.

For example, the Necker cube, a famous ambiguous figure, can appear to flip between two orientations. Classical physics would suggest a fixed perception, but quantum-inspired models allow the system to “choose” one perspective over time. Ivan Maksymov in Australia developed a quantum-AI hybrid to simulate both the Necker cube and the Rubin vase, where a vase can also appear as two faces. The AI switched between interpretations like a human, with similar timing.

Maksymov clarifies that this doesn’t mean our brains are quantum; rather, quantum models can better capture certain aspects of decision-making, such as how the brain resolves ambiguity.

Such AI systems could also help us understand how perception changes in unusual environments. Astronauts on the International Space Station experience optical illusions differently. For instance, the Necker cube tends to favor one orientation on Earth, but in orbit, astronauts see both orientations equally. This may be because gravity helps our brains judge depth something that changes in free fall.

With the Universe holding so many wonders, astronauts and the rest of us will be glad to know there are ways to study when our eyes can be trusted.