Even if there is no fur in the frame, you can easily see that a photo of a hairless Sphynx cat is depicting a cat. You wouldn’t mistake it for an elephant.
But many will be artificial intelligence vision systems. Why? Because when AI systems learn to classify objects, they often rely on visual cues—eg surface texture or normal patterns in pixels. This tendency puts them at risk of being confused by small changes that have little effect on people’s perception.
A visual system more closely aligned with human perception—one that probably emphasizes shape, for example—might confuse a cat for a mammal still the same size as a tiger; But it is unlikely to indicate an elephant.
The kinds of mistakes an AI makes reveal how it organizes visual information, potential limitations that become worrisome in high-stakes settings.
Imagine an autonomous car approaching a broken stop sign. While a human driver recognizes the sign from its shape and context, an AI that relies on pixel patterns can misclassify it, pushing the changed sign. Totally out of the “sign” category and in a different group of images that it identifies as similar, such as a billboard, advertisement or other roadside object
Together, these issues point to a disconnect between how humans perceive the visual world and how AI represents it.
We are visual perception expertsAnd We work at the crossroads of people and machine perception. Humans organize visual input into objects, meanings, and relationships shaped by experience and context. AI models do not organize visual information in the same way. This key difference explains why AI sometimes fails in surprising ways.
See objects, not properties
Imagine that there is a small, opaque object in front of you that has both straight and curved edges. But you don’t see those features; You just look at your coffee mug.
Sight is not a camera, passively recording the world. Instead, your brain quickly transforms the light your eyes absorb into objects you recognize and understand, organizing experience. Structured mental representations.
Researchers can understand how these representations are formed by examining how people perceive them Judge the similarities. Your coffee mug is not like your computer, but it is like a glass of water, despite the difference in appearance. That judgment reflects how the mug is mentally represented: not just in terms of appearance, but what the mug is used for and how it fits into daily activities.
Importantly, the mental organization of representations is flexible. Which aspects of an object change? context and goals. When packing a moving box, shape and size are most important, so your mug can be placed where it fits. But when it’s placed in the cupboard it moves next to other drinking vessels. The mug has not changed, only the way it is organized in your mind.
Human visual perception is adaptive, driven by meaning and tied to how we interact with the world.
Aligning AI with Humans
But AI systems organize visual input in a fundamentally different way than humans—not because they’re machines, but because of how narrowly they’re trained. When an AI is trained to classify a cat or an elephant, it only needs to learn which visual patterns lead to the correct label, not how those animals relate to each other or fit into the wider world.
In contrast, people learn within a broader context. When we learn what an elephant is, we can weave that representation into the tapestry of what we’ve learned: animals, shapes, habitats, and more. Since AI is only graded on label accuracy, it can rely on shortcuts that work in training but sometimes fail in the real world.
issue of Representative alignment AI refers to whether it organizes information in a way that humans do. Not to be confused with it Value alignmentWhich refers to the challenge of ensuring that AI systems follow human intended outcomes and goals.
Because human learning embeds new information into networks of prior knowledge, relationships between new and existing concepts can be studied and measured. This means that representational alignment can be a solvable problem and a step toward addressing the larger alignment challenge.
One approach to representational alignment focuses on creating AI systems that behave like humans in psychological tasks, allowing researchers to directly compare representations. For example, if people judge a cat to be more similar to a dog than an elephant, the goal is to create AI models that reach that same judgment.
one promising strategy involved Training AI on human similarity judgments collected in the lab. In this study, human participants could be shown three images and asked which two objects were more similar; For example, whether a mug is like a glass or a bowl. Incorporating this data during training encourages AI systems to learn how objects are related to each other, creating representations that better reflect how people understand the world.
Alignment out of sight
Representational alignment matters Outside of the vision systemand AI researchers Notice taken. As AI increasingly supports high-stakes decisions, differences between how machines and humans represent the world will have real consequences, even when an AI system Looks very accurate. For example, if an AI analyzing medical images learns to associate an image source or repeated image artifacts with a disease rather than the actual visual symptoms of the disease, this is clearly problematic.
AI doesn’t necessarily need to process information the way humans think, but training AI using principles derived from human perception and knowledge — such as similarity, context, and relational structure — can lead to safer, more accurate, and more ethical systems.
Erin RobbinsAssistant Professor of Psychology, University of Richmond; Eben W. DaggettAffiliated Faculty of Psychology, New Mexico State UniversityAnd Michael WoodAssociate Dean for Research and Professor of Psychology, New Mexico State University
Reprinted from this article the conversation Under Creative Commons license. read on Main article.
