A paper-and-pencil questionnaire outperforms GPT for measuring cognitive vulnerability to depression and predicting depressive symptoms.

We tested if GPT could predict changes in depressive symptoms using participants' (n = 930) causal explanations for negative life events. Results showed that 2 of 30 GPT prompts yielded output that could reliably predict changes in future depressive symptoms; but this output was not a better predictor than the traditional paper-and-pencil measure of cognitive risk for depression (Cognitive Style Questionnaire). These findings highlight potential limitations of large language models like GPT. Human thought is complex, and language may not accurately represent people's internal cognitive processes. In this case, participants' written explanations for negative life events did not contain meaningful information that could be used for differentiation (or was indicative of some latent construct). We found that people could generate equally negative causal explanations for negative events yet hold different beliefs about the changeability of those causes. Our results support the hypothesis that it is the perceived changeability, not the overall negativity, of causal beliefs that determines risk for depressive symptoms. GPT cannot yet discern this changeability as well as a paper-and-pencil questionnaire.