Even with no fur in frame, you can easily see that a photo of a hairless Sphynx cat depicts a cat. You wouldn’t mistake it for an elephant. But many artificial intelligence vision systems would. Why? Because when AI systems learn to categorize objects, they often rely on visual cues—like surface texture or simple patterns in pixels. This tendency makes them vulnerable to getting confused by small changes that have little effect on human perception.
A vision system aligned more closely with human perception—one that perhaps emphasizes shape—might still confuse the cat for a tiger, but it is unlikely to indicate an elephant. The mistakes an AI makes reveal how it organises visual information, with potential limitations that become concerning in high-stakes settings. Imagine an autonomous vehicle approaching a vandalised stop sign. While a human recognizes the sign from its shape and context, an AI relying on pixel patterns may misclassify it as a billboard or advertisement.
As experts in visual perception, we see a fundamental misalignment. People organize visual input into objects, meaning, and relationships shaped by experience. AI models don’t. Imagine a coffee mug; you don't just see edges, you see a functional object. Your brain rapidly turns light into structured mental representations. Human perception is adaptive; if packing a box, the mug’s size matters most, but in a cupboard, it belongs with other drinkware. The mug hasn’t changed—only your mental organisation.
AI systems organize input differently because of how narrowly they are trained. When an AI learns "cat" or "elephant," it only seeks visual patterns that lead to the correct label. It doesn't learn how those animals relate to the broader world. In contrast, humans weave representations into a tapestry of prior knowledge: habitats, size, and biology. Because AI is graded only on label accuracy, it relies on shortcuts that fail in the real world.
This is a problem of representational alignment: whether AI organises information in ways that resemble how people do. This is distinct from value alignment, which ensures AI pursues intended goals. Because human learning embeds information into a web of concepts, representational alignment may be a solvable problem. One approach focuses on building AI that behaves like humans on psychological tasks. For example, if people judge a cat as more similar to a dog than to an elephant, the goal is to build models that arrive at those same judgments.
One promising technique involves training AI on human similarity judgments collected in the lab. Including this data encourages systems to learn how objects relate, producing representations that better reflect human understanding. This matters beyond vision. If an AI analyzing medical images learns to associate image artifacts with disease rather than the actual pathology, the consequences are grave. AI doesn’t necessarily need to process information exactly like people, but training it using principles of human cognition — similarity, context, and relational structure — can lead to safer, more ethical systems.
The Conversation