Exploring the Link Between the Emotional Recall Task and Mental Health in Humans and LLMs

The ability of large language models to recall human emotions provides a novel opportunity to investigate links among memory, affect, and mental health. This study explores whether the Emotional Recall Task (ERT), a free word-association paradigm, can reveal cognitive markers of distress in both humans and large language models (LLMs). Using spreading activation simulations grounded in cognitive network science, we examined how the recall of emotional concepts (e.g., stress, anxiety, and depression) relates to psychometric measures of well-being and personality. In Study 1, correlations were tested between activation dynamics and clinical scales (DASS-21, PANAS, and Life Satisfaction) in human participants (N = 1200) and artificial participants generated by GPT-4, Claude Haiku, and Anthropic Opus. For both human and LLM samples, spreading activation was modeled from participants’ ERT words within a human-derived semantic network, enabling a direct comparison of structural activation dynamics rather than psychological states. Humans with higher distress scores exhibited stronger, faster, and more persistent activation of negative concepts, supporting theories of rumination and memory bias. GPT-4 approximated human-like trajectories most closely, though with reduced variability. Study 2 linked recall dynamics with the Big Five traits, confirming that neuroticism predicted greater activation of negative concepts, while extraversion acted as a protective factor. While LLMs lack autobiographical memory, their semantic activation partially mirrored human associations. These findings demonstrate that network-based spreading activation analysis can reveal cognitive signatures of distress while also highlighting the limits of LLMs in modeling human affect.