A recent study has shown that artificial intelligence and people actually see objects in different ways. Where humans tend to focus on what things mean, AI mostly pays attention to how objects look, their shape, color, or surface details seem to guide the AI’s attention.
The research, carried out by scientists at the Max Planck Institute for Human Cognitive and Brain Sciences, found that even when AI seems to make decisions like humans, the way it gets there can be completely different. This tendency, described as a "visual bias" in AI, means machines put much more weight on visual features than on what things actually are. That difference matters because it affects how much we can rely on AI when tasks appear to need human-like understanding.
To explore this, the team set up a direct comparison between humans and deep neural networks by using a large image judgment task. Human participants were shown groups of three images and asked to pick the one that didn’t fit, a simple way to see how people naturally group things. The same sets of images were tested on deep neural networks trained to recognize objects. By treating the networks like human participants, the researchers could dig into how both systems sort the world.
They broke down the key patterns behind these choices, calling them "dimensions." These dimensions captured both how things look and what they mean. People mostly focused on meaning, but the AI leaned heavily on surface-level visuals.
At first glance, the AI’s internal patterns seemed quite similar to those found in humans. But when the team looked closer, it became clear that many of these connections were shallow. Sometimes, the AI would group animals together with things that looked vaguely similar but had no real link. Where humans consistently kept animals in their own group, the AI often mixed in objects that simply shared a shape or color.
The team ran a series of strict tests to check whether the AI’s internal patterns genuinely reflected the properties they appeared to, or if they just looked convincing on the surface. They generated brand new images to see which ones the AI thought best matched specific patterns, removed features from images to watch how the AI’s choices changed, and created heat maps to track which parts of an image the AI actually focused on.
Although the AI did hold some meaningful groupings inside its system, these only roughly matched the way humans understand objects. Often, the AI gave answers that looked correct, but its reasoning was very different, usually based on simple image features rather than the deeper categories people use.
The researchers also checked whether these internal mismatches would show up in actual choices. By removing individual patterns from the AI and human models, they watched how the decision-making shifted. The results showed that AI’s focus on visual features played a big part in how it grouped things. Even when humans and AI picked the same object, they often got there for completely different reasons.
What made this study stand out was how it directly linked behavior and internal thinking in both systems. Previous studies often just looked at whether people and AI behaved in a similar way, but not what was actually shaping those choices. This time, the researchers opened up the systems and looked at the foundations.
The findings suggest that AI systems, even when they seem to act like humans, might still process the world in a way that's quite separate from how people think. This raises new questions about whether AI truly "understands" the categories it seems to recognize or whether it's just extremely good at picking up surface-level patterns.
The study also offers useful tools for future research. The methods used here could help build AI models that focus more on meaning, which might make them more predictable and more in tune with how people make decisions.
By shining a light on these hidden differences, the study shows that one of the big challenges in artificial intelligence is closing the gap between looking human and actually thinking like one.
Read next:
• AI Firms Face New Legal Boundaries as Court Differentiates Fair Use from Copyright Theft
• No Browser Needed: ElevenLabs Rolls Out Text-to-Speech App for Smartphones
The research, carried out by scientists at the Max Planck Institute for Human Cognitive and Brain Sciences, found that even when AI seems to make decisions like humans, the way it gets there can be completely different. This tendency, described as a "visual bias" in AI, means machines put much more weight on visual features than on what things actually are. That difference matters because it affects how much we can rely on AI when tasks appear to need human-like understanding.
To explore this, the team set up a direct comparison between humans and deep neural networks by using a large image judgment task. Human participants were shown groups of three images and asked to pick the one that didn’t fit, a simple way to see how people naturally group things. The same sets of images were tested on deep neural networks trained to recognize objects. By treating the networks like human participants, the researchers could dig into how both systems sort the world.
They broke down the key patterns behind these choices, calling them "dimensions." These dimensions captured both how things look and what they mean. People mostly focused on meaning, but the AI leaned heavily on surface-level visuals.
At first glance, the AI’s internal patterns seemed quite similar to those found in humans. But when the team looked closer, it became clear that many of these connections were shallow. Sometimes, the AI would group animals together with things that looked vaguely similar but had no real link. Where humans consistently kept animals in their own group, the AI often mixed in objects that simply shared a shape or color.
The team ran a series of strict tests to check whether the AI’s internal patterns genuinely reflected the properties they appeared to, or if they just looked convincing on the surface. They generated brand new images to see which ones the AI thought best matched specific patterns, removed features from images to watch how the AI’s choices changed, and created heat maps to track which parts of an image the AI actually focused on.
Although the AI did hold some meaningful groupings inside its system, these only roughly matched the way humans understand objects. Often, the AI gave answers that looked correct, but its reasoning was very different, usually based on simple image features rather than the deeper categories people use.
The researchers also checked whether these internal mismatches would show up in actual choices. By removing individual patterns from the AI and human models, they watched how the decision-making shifted. The results showed that AI’s focus on visual features played a big part in how it grouped things. Even when humans and AI picked the same object, they often got there for completely different reasons.
What made this study stand out was how it directly linked behavior and internal thinking in both systems. Previous studies often just looked at whether people and AI behaved in a similar way, but not what was actually shaping those choices. This time, the researchers opened up the systems and looked at the foundations.
The findings suggest that AI systems, even when they seem to act like humans, might still process the world in a way that's quite separate from how people think. This raises new questions about whether AI truly "understands" the categories it seems to recognize or whether it's just extremely good at picking up surface-level patterns.
The study also offers useful tools for future research. The methods used here could help build AI models that focus more on meaning, which might make them more predictable and more in tune with how people make decisions.
By shining a light on these hidden differences, the study shows that one of the big challenges in artificial intelligence is closing the gap between looking human and actually thinking like one.
Read next:
• AI Firms Face New Legal Boundaries as Court Differentiates Fair Use from Copyright Theft
• No Browser Needed: ElevenLabs Rolls Out Text-to-Speech App for Smartphones
