Abstract
Given differences in the associative memory architectures of humans and AI, it is not clear whether AI generates and interprets visual-linguistic associations the way humans do. To investigate this, human participants and two large language models (LLMs), ChatGPT and Gemini, were asked to generate novel words and corresponding images that form meaningfully linked dyads. The LLMs and a second group of human participants were presented with shuffled sets of these words and images and asked to rematch them in ways they felt were most meaningful. Humans performed similarly regardless of whether the pairs were human or LLM-generated, while LLMs were better at matching LLM-generated pairs. Overall, LLMs outperformed humans at recovering the intended pairings. These findings suggest that despite architectural differences, the way in which AI generates and detects associations between language and visual stimuli is not only like that of humans, but they may be superior at recognizing cross-domain associations that humans find meaningful.