New research explores how language-based AI understands color metaphors
ChatGPT, like other AI language models, is trained by analyzing billions of words.
It learns how words go together and what people mean when they say things like “seeing red” or “feeling blue.”
But unlike humans, ChatGPT has never actually seen a red apple or a blue sky. So, can it truly understand color in the same way people do?
That’s the big question behind a new study published in Cognitive Science.
Led by Professor Lisa Aziz-Zadeh of USC and a team of researchers from universities around the world and Google DeepMind, the study explores how different types of people—and an AI—interpret color-related language.
Aziz-Zadeh’s lab focuses on how the brain supports complex human behaviors like emotion, language, and social understanding.
This research zoomed in on how experience shapes our understanding of color words, especially metaphors like “on red alert” or more unusual phrases like “a very pink party.”
The study included four groups: people with typical color vision, people who are color-blind, painters who work with colors professionally, and ChatGPT. All groups were asked to match colors with abstract words like “physics,” interpret familiar color metaphors, and explain more unusual ones.
One surprising result was that color-blind people and those with normal vision gave very similar answers. This suggests that you don’t have to see color to understand its symbolic or emotional meaning.
Painters, on the other hand, did better than everyone else when it came to understanding unfamiliar color metaphors. This shows that regular hands-on experience with color gives people a richer, deeper understanding of what color can mean in language.
ChatGPT also gave consistent and thoughtful answers. For example, when asked about a “very pink party,” it explained that pink is often linked to “happiness, love, and kindness,” so the party likely had good vibes. These responses show that the AI has learned a lot about cultural and emotional associations from the text it has read.
But ChatGPT still had some trouble, especially when it was asked to interpret new, unfamiliar color metaphors like “the meeting made him burgundy” or to explain the opposite of certain color meanings. It also didn’t talk about physical or visual experiences as often as human participants did.
The researchers say this highlights a key difference between AI and humans. While AI can be great at finding patterns in language, it still lacks the “embodied” experiences—like painting or seeing a sunset—that shape how people understand the world.
Future AI models might one day include sensory data to help close this gap, but for now, there’s still a difference between mimicking language and truly understanding what it means.
