What is it about?
Can Large Language Models (LLMs) think about color the way people do? Our study compares several GPT versions from GPT-3 to GPT-4o with the color-word associations of more than 10,000 Japanese speakers for matching 17 colors to words from eight semantic categories. We found that LLMs show a significant and interesting gap in mapping colors to words, despite good color-discrimination skills shown in previous studies. This raises the possibility of systematic differences in semantic memory structures between humans and LLMs in representing color-word associations.
Featured Image
Photo by Pierre Blaché on Unsplash
Why is it important?
Color-word associations underpin everything from everyday communication to professional design processes, tapping into deep semantic memory structures. Assessing LLMs on this task moves beyond conventional language benchmarks—revealing where AI’s conceptual “understanding” aligns with or departs from human cognition. The findings expose systematic gaps in LLM semantic representations.
Perspectives
In some earlier studies, large language models (LLMs) were shown to handle color distinctions in ways that resemble human perception (for example, recognizing that green is closer to blue than to yellow). Our study, however, takes a step further by looking at the semantic associations of colors—that is, what kinds of words or concepts LLMs connect with particular colors (such as associating yellow with a sour taste, or red with anger). We found that with each new generation of GPT, the model’s responses are gradually moving closer to human-like patterns, yet there remain intriguing and sometimes surprising gaps. To me, these results highlight differences in the structure of semantic memory between humans and LLMs—differences that cannot be captured simply by testing color discrimination. On a personal note, this work has been a meaningful milestone for me. It has not only provided insight into how LLMs diverge from humans in subtle ways, but also sparked deeper reflection on alignment and the broader challenges of bridging human and AI cognition.
Dr Makoto Fukushima
Read the Original
This page is a summary of: Advancements and limitations of LLMs in replicating human color-word associations, Discover Artificial Intelligence, May 2025, Springer Science + Business Media,
DOI: 10.1007/s44163-025-00323-8.
You can read the full text:
Contributors
The following have contributed to this page
