Explaining GPTs' schema of depression: A machine behavior analysis
Ganesan, A. V., Varadarajan, V., Lal, Y. K., Eijsbroek, V. C., Kjell, K., Kjell, O. N. E., Dhanasekaran, T., Stade, E. C., Eichstaedt, J. C., Boyd, R. L., Schwartz, H. A., & Flek, L. (2026). Explaining GPTs' schema of depression: A machine behavior analysis. Journal of Psychopathology and Clinical Science. https://doi.org/10.1037/abn0001144
@article{ganesan2026explaining,
title = {{Explaining GPTs' schema of depression: A machine behavior analysis}},
author = {Adithya V. Ganesan and Vasudha Varadarajan and Yash Kumar Lal and Veerle C. Eijsbroek and Katarina Kjell and Oscar N.E. Kjell and Tanuja Dhanasekaran and Elizabeth C. Stade and Johannes C. Eichstaedt and Ryan L. Boyd and H. Andrew Schwartz and Lucie Flek},
year = {2026},
journal = {Journal of Psychopathology and Clinical Science},
doi = {10.1037/abn0001144},
url = {https://doi.org/10.1037/abn0001144}
}
Abstract
The use of large language models (LLMs) such as ChatGPT (GPT-4/GPT-5) for mental health is growing rapidly, including their use to assess and help people with mood disorders, such as depression. However, we have a limited understanding of the schema of mental disorders in these language models, that is, how they internally associate and interpret the symptoms. In this work, we used contemporary measurement theory to decode how GPT-4 and GPT-5 interrelate depressive symptoms, providing an explanation of how large language models behave in clinical applications. First, we evaluated whether these models can reliably detect depressive constructs, finding that GPT-4 (a) demonstrated strong convergent validity with standard instruments and expert judgments (r = .70-.81). We mapped the model's internal schema of depression, showing that it (b) organized symptoms into an interconnected network (symptom correlations r = .23-.78) that largely aligned with the established clinical literature. However, it (c) underemphasized the relationship between suicidality and other symptoms while overemphasizing the importance of psychomotor symptoms. Finally, the associations (d) suggested novel hypotheses for symptom mechanisms, for instance, indicating that sleep and fatigue are broadly influenced by other depressive symptoms, whereas worthlessness/guilt is selectively associated with depressed mood. GPT-5 showed slightly lower convergence with self-report, which our analysis attributed to changes in the relationships among symptoms. These insights provide a generalizable, empirical foundation for understanding language models' mental health assessments across models and disorders.
