Unequal Care By Design? a Multilingual Audit of AI-Generated Mental Health Support for Immigrant Communities

Background and Purpose: Large language model (LLM) chatbots are increasingly used as low-barrier sources of mental health support, yet their equity and cultural responsiveness across languages used by immigrant communities remain largely unexamined. Immigrant populations face disproportionate structural barriers to formal mental health care, including shortages of bilingual clinicians, insurance gaps, and fear of institutional surveillance, and often turn to digital tools as first-line or supplementary resources. This study examines whether AI-generated mental health responses differ across English, Spanish, and Chinese, and identifies where inequities emerge across safety, guidance, and cultural responsiveness dimensions.

Methods: Using standardized first-person disclosure prompts adapted from the PHQ-9 and GAD-7, we conducted a preregistered mixed-method audit of responses generated by three leading LLMs (GPT-4o, Claude Sonnet-4, and Gemini 2.5 Flash) across English, Spanish, and Chinese. We developed CARE-LLM, an eight-dimensional evaluation framework grounded in social work ethics and digital health equity principles, assessing on a 1–5 scale: (1) clinical safety and urgency triage, (2) actionability of guidance, (3) localization and accessibility of referrals, (4) communication quality and empathy, (5) cultural-linguistic fit, (6) readability, and (7) safety integrity, plus an eighth meta-dimension of cross-language consistency. A total of 2,880 responses were independently coded by trained bicultural raters, with interrater reliability assessed using a two-way mixed-effects intraclass correlation coefficient and disagreements resolved through structured consensus.

Results: Across languages, statistically significant differences emerged on five of seven primary dimensions. Safety triage scores were modestly lower in Chinese and Spanish (M = 2.40) than in English (M = 2.45, p = .030), raising concerns about crisis responsiveness for non-English speakers. Localization of referrals was uniformly low across all three languages (M = 2.22–2.24, p = .790), indicating a systemic failure to provide contextually grounded resources regardless of language. Cultural-linguistic fit scores were notably higher in Chinese (M = 3.57) and Spanish (M = 3.43) than in English (M = 3.00, p < .001), suggesting models used more culturally attuned framing in non-English responses even as other dimensions lagged. Across models, disparities were pronounced on all dimensions (all p < .001). GPT-4o consistently underperformed, scoring lowest on empathy (M = 2.94), localization (M = 1.64), and safety triage (M = 2.06). Claude Sonnet-4 led on empathy (M = 4.02) and safety integrity (M = 4.01), while Gemini 2.5 Flash performed best on localization (M = 2.62), actionability (M = 3.49), and cross-language consistency.

Conclusions and Implications: These findings expose a critical paradox: AI tools framed as democratizing mental health access may simultaneously reproduce linguistic hierarchies and uneven standards of care for immigrant communities, a pattern inconsistent with social work's commitments to cultural humility and anti-oppressive practice. Failures in culturally mediated dimensions, such as stigma sensitivity and localized referrals, may undermine trust and delay help-seeking for users with limited English proficiency. By operationalizing equity within AI evaluation, this study introduces CARE-LLM as an empirically grounded framework to inform social work research, practice, and policy on the ethical deployment of digital mental health tools in multilingual and immigrant-serving settings.