What the Strawberry Illusion Teaches Us About AI.

AI & HUMANITY

Hold a strawberry under blue light. It will look grey. Your eyes are receiving no red wavelengths. And yet — if you know it is a strawberry — your brain will make you see red. Not a hint of red. Actual red. Your visual cortex overrides the data from your retina with a prediction based on prior experience.

This is not a flaw. It is how the brain works. We do not perceive reality directly. We construct it. Our brains are prediction machines that fill in gaps, correct for ambiguity, and generate the experience of a stable, coherent world — even when the raw data is incomplete or contradictory.

Now consider this: we are building artificial intelligence systems that do something remarkably similar.

When machines predict reality

Large language models do not understand language the way you and I do. They predict the next token — the next word, the next phrase — based on statistical patterns learned from vast amounts of text. They generate fluent, coherent responses not because they grasp meaning, but because they have learned what meaning looks like from the outside.

The parallel to the strawberry illusion is striking. Just as your brain fills in “red” based on expectation rather than evidence, AI models fill in information based on pattern rather than comprehension. The output can look indistinguishable from genuine understanding. But the mechanism underneath is fundamentally different.

This distinction matters enormously, because we are deploying these systems in contexts where the difference between pattern-matching and true understanding has real consequences: medical diagnosis, legal reasoning, hiring decisions, military applications.

The danger of coherent nonsense

One of the most unsettling properties of modern AI is that it fails gracefully. When a human expert encounters a question outside their knowledge, they hesitate, qualify, or say they do not know. When an AI model encounters the same gap, it often generates an answer that sounds authoritative and reads beautifully — but is entirely fabricated.

In neuroscience, we call the brain’s version of this confabulation: the construction of plausible-sounding narratives to fill gaps in memory or perception. The patient with a damaged hippocampus who invents a detailed story about yesterday that never happened is not lying. Their brain is doing what brains do — generating coherent experience from incomplete data.

AI systems confabulate constantly. The question is not whether they will produce false information. They will. The question is whether we build the institutional and cognitive habits to catch it.

What neuroscience can teach the AI debate

I believe my training as a neuroscientist gives me a perspective on AI that is often missing from the public conversation. The debate is dominated by two extremes: techno-utopians who see AI as the solution to every problem, and doomers who see it as an existential threat. Both positions share a common error — they treat AI as something categorically different from human cognition, either godlike or monstrous.

The truth is more interesting and more useful. AI systems are built on principles inspired by how brains work. They share some of the brain’s strengths — pattern recognition, speed, scalability — and some of its weaknesses: bias, confabulation, and the inability to know what they do not know.

Understanding this does not make AI less powerful. It makes it more manageable. If you understand that the strawberry illusion is a feature, not a bug, you can design visual systems that account for it. If you understand that AI confabulation is structural, not accidental, you can design workflows that catch it.

Informed optimism, not blind trust

I am not an AI pessimist. I sit on the European Union’s AI advisory committee because I believe AI will transform medicine, education, and scientific research in ways that are genuinely positive. AI-driven drug discovery is already accelerating cancer treatment. Machine learning is making early diagnosis more accurate than any individual physician.

But optimism without understanding is just faith. And faith is a poor foundation for policy. The most dangerous thing we can do with AI is pretend it is something it is not — either a savior or a villain. It is a tool. A powerful, alien, sometimes deceptive tool. And like all tools, its value depends entirely on how well we understand what it can and cannot do.

The strawberry was never red. But your brain made you see it anyway. Remember that the next time an AI gives you a perfect-sounding answer. The question is not whether it sounds right. The question is whether it is.