When AI Isn’t Enough: Preserving Meaning through Human and Community Verification In Qualitative HIV Prevention Research

Background and Purpose: Artificial intelligence (AI) offers increased efficiency in qualitative research, particularly for transcription and data organization. However, its ability to accurately represent participant meaning remains unclear. These concerns are especially important in research with marginalized populations, where language reflects cultural context and lived experience. For Black and Latino sexual minority men (SMM), forms of expression such as African American Vernacular English (AAVE), Spanglish, and code-switching may be underrepresented in AI training data, increasing the risk of misinterpretation. In HIV prevention research, where medical mistrust is a primary barrier to care, misinterpretation of culturally specific language may obscure aspects of lived experience. Aligned with the SSWR 2027 theme of advancing inquiry for the common good, this study examines whether AI-assisted transcription preserves meaning in equity-focused qualitative data and evaluates how human and community verification can improve interpretive accuracy and inform the refinement of medical mistrust measurement.

Methods: We conducted a secondary qualitative analysis of audio recordings from two 8-session CBT-based HIV prevention interventions: Siempre Seguiré (Spanish-language for Latino SMM) and Still Climbin’ (English-language for Black SMM). Sessions were purposively selected based on in-depth discussions of medical mistrust. Audio files were transcribed using AI-assisted software to generate initial transcripts. These transcripts were then systematically reviewed against the original audio by a bicultural research team, who corrected errors and restored culturally specific meanings to produce verified transcripts for analysis. The team met regularly to ensure consistent interpretation across transcripts. Verified transcripts were coded using thematic analysis in Dedoose to identify domains of medical mistrust. Coded themes were compared to a domain-level item matrix derived from existing medical mistrust measures to identify conceptual gaps and context-specific phrasing needs. Findings were reviewed with Community Advisory Committees and Community Advisory Boards (CACs/CABs) to ensure linguistic and cultural validity.

Results: AI-assisted transcription improved efficiency but also introduced errors and misinterpretations, particularly in culturally specific language. During transcript verification, the research team identified instances where language describing ongoing or repeated experiences was inaccurately represented, which could affect interpretation. In Spanish-language sessions, some culturally grounded expressions were translated into more standardized phrasing that reduced their original meaning and intensity. Without correction, these issues would have affected how themes were interpreted. Human and community verification were important for restoring meaning and preserving linguistic nuance. Findings from the verified transcripts reflected commonly identified domains of medical mistrust, while also highlighting areas where current approaches may not fully capture participants’ experiences, including issues related to immigration status and medical privacy. Community input supported the refinement of more culturally responsive language to better reflect lived experiences of mistrust.

Conclusions and Implications: Findings suggest that AI alone is insufficient for qualitative research in equity-focused contexts. Reliance on automated processes may misrepresent culturally grounded experiences and mask structural inequities. Integrating human and community verification improves interpretive accuracy and strengthens the cultural validity of qualitative findings and measurement development. This study provides practical guidance for the responsible use of AI in social work research and advances methods for studying medical mistrust in HIV prevention.