Abstract
Well-being is commonly defined in terms of comfort, happiness, functioning, and flourishing. Scholars distinguish between subjective well-being (i.e., perceiving life as pleasant) and psychological well-being (i.e., perceiving life as meaningful). While advances in natural language processing have enabled automated assessments of subjective well-being from language, their ability to capture psychological well-being remains underexplored. Across three studies (one preregistered), we examined how well language-based models predict self-reported subjective and psychological well-being. Participants provided verbal or written responses about their satisfaction with life and autonomy, along with standard questionnaire measures. We used contextual word embeddings from transformer-based models to predict well-being scores. Language-based predictions correlated moderately with questionnaire measures of both constructs (rs = .16-.63) and generalized across well-being domains (rs = .15-.50), though these associations were weaker than previously work (rs = .72-.85). Autonomy was consistently less predictable than satisfaction with life. Comparisons with GPT-3.5 and GPT-4 revealed that both models outperformed BERT in predicting satisfaction with life (r = .71 and .75) and modestly improved predictions of autonomy (r(GPT‑4) = .49). Supervised dimension projections revealed that satisfaction with life responses clustered around positive emotion and social themes, whereas autonomy responses showed more individualized linguistic patterns. These findings suggest that language-based tools are well-suited for assessing hedonic well-being but face challenges with more abstract, eudaimonic constructs. Future research should refine modeling approaches to enhance the detection of complex psychological states while striking a balance between interpretability, accuracy, and usability.