AI may be able to 'taste' colours: study finds

The research explored the phenomenon of crossmodal correspondences—the human tendency to associate specific shapes or colours with particular tastes.

The research explored the phenomenon of crossmodal correspondences—the human tendency to associate specific shapes or colours with particular tastes.

A recent study suggests that artificial intelligence (AI) models, like OpenAI's ChatGPT, might be capable of 'tasting' colours and shapes in ways similar to how humans do.

The research, titled 'Crossmodal Correspondences in Language Models: Can AI 'Taste' Colours and Shapes?', explored the phenomenon of crossmodal correspondences—the human tendency to associate specific shapes or colours with particular tastes. For example, people often link round shapes with sweetness or umami, and sharp angles with bitterness. But can AI make these associations too?

The study, authored by Kosuke Motoki from The University of Tokyo, Charles Spence from the University of Oxford, and Carlos Velasco from BI Norwegian Business School, was recently published in Cognition, an international journal that publishes experimental research on the study of the human mind.

The researchers tested two versions of ChatGPT (3.5 and 4) on crossmodal associations across three languages: English, Japanese, and Spanish. The findings revealed striking similarities to human patterns.

In terms of shape-taste associations, ChatGPT consistently linked round shapes with sweet and umami tastes, while angular shapes were associated with sour, salty, and bitter tastes—echoing typical human responses.

Regarding colour-taste connections, ChatGPT-3.5 had difficulty, but ChatGPT-4 demonstrated a better alignment with human patterns, associating pink with sweetness, yellow with sourness, and black with bitterness.

The study also observed that the language used in the prompts influenced AI's responses. Notably, Japanese prompts produced more varied associations, especially with umami flavours.

ChatGPT-4 outperformed its predecessor, ChatGPT-3.5, particularly in English and Spanish. "While AI doesn't taste or smell like humans, it can form associations based on patterns in language data," the study explains.

Although AI doesn't physically experience taste or smell, the study suggests it learns to replicate sensory associations through the vast amounts of data it processes.

The researchers note that these findings could have significant practical applications, such as in marketing and product design, where understanding sensory connections is crucial.

As AI technology continues to evolve, its ability to make sensory associations is likely to improve, offering new possibilities in various fields, from consumer goods to user experience design.