Can a machine that’s never seen color understand what it means to feel blue or go red with rage?
A new study in Cognitive Science suggests not quite. Researchers compared how sighted and colorblind adults, visual artists, and ChatGPT interpret color metaphors. The findings reveal just how much human understanding still hinges on embodied experience, even when language seems to be enough.
While AI models like ChatGPT can generate plausible color associations, they often miss the mark when asked to explain or manipulate metaphorical uses of color. The study shows that firsthand experience with color, especially through visual art, gives people a deeper grasp of abstract language than text-trained models can match.
Testing Human and Machine Minds
The research, led by neuroscientist Lisa Aziz-Zadeh of the University of Southern California, asked four groups to tackle the same challenge: assign colors to abstract concepts and interpret both familiar and unfamiliar color metaphors.
Participants included:
- Colorseeing adults
- Colorblind adults
- Professional painters
- ChatGPT (a popular large language model)
Each group was prompted to link colors to non-obvious words like “physics” and explain figurative phrases such as “on red alert” and “it was a very pink party.”
Colorblind but Not Concept-Blind
One striking finding was how similar the responses were between colorseeing and colorblind adults. Despite limited or no color perception, colorblind participants still made consistent, coherent metaphorical associations. This suggests that language alone carries powerful color-based connotations, even for those with little visual input.
“This challenges the idea that visual experience is necessary to build metaphorical meaning around color,” said the study’s authors.
Embodied Expertise: Painters Excel
But painters (people who regularly work with pigments) stood out. They were significantly better at interpreting novel metaphors, showing that tactile, embodied experience seems to enrich conceptual understanding of color. In other words, mixing paints may help people mix metaphors more fluently too.
ChatGPT Gets Close, But Misses Key Nuance
ChatGPT’s performance was more uneven. The model could produce consistent associations and often cited emotional or cultural reasons for its color choices. For example, it explained the phrase “very pink party” by saying, “Pink is often associated with happiness, love, and kindness,” implying a party filled with “positive emotions and good vibes.”
Yet ChatGPT faltered when asked to reverse color metaphors or interpret unusual ones. The AI often lacked coherent justifications for its responses, particularly when metaphorical meanings became more abstract or ambiguous.
What Language Models Still Can’t Do
According to Aziz-Zadeh, this exposes a core gap between human and machine cognition.
“This project shows that there’s still a difference between mimicking semantic patterns and the spectrum of human capacity for drawing upon embodied, hands-on experiences in our reasoning,” she said.
Unlike humans, AI lacks direct sensory input. It has never seen a sunset or mixed paint on a canvas. So while it can infer associations from language, it cannot feel them.
Where Does This Leave AI?
The study’s implications go beyond color. It raises broader questions about how much of human understanding can truly be captured through language alone. Could AI come closer to human cognition by integrating sensory data such as images, sounds, or tactile input? Or will there always be a gap between simulation and experience?
As LLMs grow more advanced, this research acts as a reminder: reading about red is not the same as seeing it. And language, no matter how rich, may never fully replace lived experience.
Journal Reference
Nadler, E. O., Guilbeault, D., Ringold, S. M., Williamson, T. R., Bellemare-Pepin, A., Comșa, I. M., Jerbi, K., Narayanan, S., & Aziz-Zadeh, L. (2025). Statistical or Embodied? Comparing Colorseeing, Colorblind, Painters, and Large Language Models in Their Processing of Color Metaphors. Cognitive Science. https://doi.org/10.1111/cogs.70083
ScienceBlog.com has no paywalls, no sponsored content, and no agenda beyond getting the science right. Every story here is written to inform, not to impress an advertiser or push a point of view.
Good science journalism takes time — reading the papers, checking the claims, finding researchers who can put findings in context. We do that work because we think it matters.
If you find this site useful, consider supporting it with a donation. Even a few dollars a month helps keep the coverage independent and free for everyone.
“Yet ChatGPT faltered when asked to reverse color metaphors or interpret unusual ones” sure it would. It can only do what it has been trained to do.