Recent research has explored how ChatGPT, an advanced language model by OpenAI, interprets color metaphors, a common feature in the English language. A study published in *Cognitive Science* in July 2025, led by Professor Lisa Aziz-Zadeh, compared ChatGPT's responses to those of color-seeing adults, color-blind adults, and painters. Participants assigned colors to abstract words and interpreted familiar and novel color metaphors.
The findings revealed that ChatGPT generated consistent color associations, often referencing emotional and cultural associations. For example, interpreting "a very pink party," ChatGPT noted that "pink is often associated with happiness, love, and kindness, suggesting that the party was filled with positive emotions and good vibes." However, the model struggled with novel metaphors and inverted color associations.
Another study investigated ChatGPT's ability to replicate human-like color-concept associations. Researchers assessed the model's capacity to estimate human color-concept associations without extra training. They found that ChatGPT's ratings correlated with human ratings, indicating that language models can learn color-concept associations from language data.
These studies highlight the progress and limitations of AI in understanding color metaphors. While ChatGPT generates color associations based on linguistic data, it lacks the sensory experiences that inform human understanding. This leads to challenges in interpreting novel or inverted metaphors. Future research may focus on integrating sensory input into AI models to bridge this gap.