Towards social superintelligence? AI infers diverse psychological traits from text without specific training, outperforming human judges

Large Language Models (LLMs) have recently demonstrated impressive capabilities in domains requiring higher-order cognition. This study investigates whether LLMs can also perform a core social-cognitive function: forming predictive models of individuals’ psychological traits from minimal input (“trait inference”). Extending earlier work that has focused almost exclusively on Big Five personality factors, we asked GPT-4 to anticipate responses on a battery of 30 validated scales spanning personality, affect, values, and interpersonal style. A total of 1,011 participants wrote short self-descriptive texts and completed the questionnaires. An LLM was tasked with predicting participants’ questionnaire responses solely from their self-descriptions, without any task-specific training. Human judges attempted the same task, providing a direct benchmark. The LLM’s predictions correlated with participants’ self-reports (r = .35; disattenuated r = .41)—accuracy comparable to that typically observed among real-world friends and substantially higher than that of human judges (r = .20; disattenuated r = .23). Across scales, the performance of the LLM and human judges was moderately correlated. These findings highlight LLMs’ emerging capacity for sophisticated social inference, opening new avenues for computational psychology while raising important ethical concerns about large-scale psychological profiling. • AI infers 30 diverse psychological traits from text with no task-specific training • Zero-shot GPT-4 scores r = .41 vs human r = .23 on 1,000 autobiographies • Accuracy rivals close-friend ratings on personality and coping constructs • Trait prediction is strongest for personality/coping; weaker for values & sociopolitical attitudes. • Findings raise ethical questions for mental health, education and surveillance.