Why AI Still Can’t Really Feel Blue or See Red

Can machines of never-before-seen colors understand what it means to turn red with blue or anger?

A new study Cognitive Science The suggestion is not complete. Researchers compared how adults with vision and color blindness, visual artists and Chatgpt interpreted color metaphors. These findings suggest that even if language seems sufficient, human understanding still depends on the experience embodied.

While AI models like chatgpt can produce reasonable color associations, they often miss marks when asked to explain or manipulate the metaphorical uses of colors. The study shows that first-hand color experience, especially through visual art, enables people to master abstract languages than text-trained models to match.

Testing the thinking of people and machines

The study, led by University of Southern California neuroscientist Lisa Aziz-Zadeh, asked four groups to deal with the same challenge: assigning colors to abstract concepts and explaining familiar and strange color metaphors.

Participants include:

• Adults
• Color blind adults
• Professional painter
• chatgpt (a popular big language model)

Each group was prompted to link colors with non-obvious words like “physics” to symbolic phrases such as “red alert” and “it’s a very pink party.”

Color blindness but not concept blindness

One amazing discovery is how similar the responses are between color and color blind adults. Despite limited or no color perception, color blind participants remained consistent, coherent metaphorical associations. This suggests that language alone has a strong color-based connotation even for those with very little visual input.

“This challenges the notion that visual experience is necessary for building metaphorical meaning around color,” said the authors of the study.

Demonstrate expertise: The painter performs well

But painters (those who often use pigments) stand out. They performed well in explaining novel metaphors, showing tactile sensations, and specific experiences seemed to enrich the conceptual understanding of color. In other words, mixing paints may help people mix metaphors more fluently.

Chatgpt approaches, but misses key nuances

Chatgpt’s performance is even more unbalanced. The model can produce consistent associations and often cite emotional or cultural reasons for its color choice. For example, it explains a “very pink party” by saying “very pink party” that “pink is often associated with happiness, love and kindness”, which means a political party full of “positive emotions and good resonance.”

However, Chatgpt falters when asked to reverse the color metaphor or explain the abnormality. AI often lacks coherent reasons for its response, especially when metaphorical meanings become more abstract or ambiguous.

Which language models still cannot be done

According to Aziz-Zadeh, this reveals the core gap between human and machine cognition.

“The project shows that there is still a difference between imitation semantic patterns and human abilities that draw on the embodiment of our reasoning, and that hands-on experience remains,” she said.

Unlike humans, AI lacks direct sensory input. It has never seen sunsets or mixed paint on canvas. So while it can infer correlations from language, it cannot feel them.

Where will this leave AI?

The meaning of this study goes beyond color. It raises a broader question about human understanding that is truly captured through language alone. Can AI get closer to human cognition by integrating sensory data such as images, sounds, or haptic inputs? Or is there always a gap between simulation and experience?

As LLM grows higher and higher, this study reminds you: reading about red is different from seeing it. Language, no matter how rich it is, may never completely replace life experience.

Journal Reference

Nadler, EO, Guilbeault, D., Ringold, SM, Williamson, TR, Bellemare-Pepin, A. Statistics or reflections? Comparison of color, color blindness, painter and big language models when dealing with color metaphors. Cognitive Science.

Related