Text or Speech, the Gap Persists: AI-based Depression Screening Is Better Calibrated for Lower Socioeconomic Status Users

Abstract

The interactional design of digital mental health systems can shape what people say, how they say it, and whose distress becomes legible to automated models. We developed a multimodal survey chatbot, M2SurveyBot with clinical psychologists, to examine whether interaction modality that is, typing versus speech, affects self-disclosure and quality of depression screening. In a study with 455 U.S. participants, typed and spoken modalities yielded statistically indistinguishable predictive performance; acoustic features alone carried near-zero signal. Critically, we find a robust socioeconomic status (SES)-related calibration gap replicated across supervised machine learning models and GPT-4o zero-shot estimation: predictions track reported depression severity more closely for lower-SES participants, suggesting socioeconomic differences in linguistic expression of distress. These findings challenge the assumption that multimodal collection produces more equitable screening, while establishing spoken interaction as a viable, accessible alternative to typed text, potentially lower-barrier for populations for whom typing poses challenges.