Conversational AI in Hereditary Cancer Care: Sociotechnical Study of Responsible Design Requirements.
Individuals with BRCA1/2 germline variants face complex, preference-sensitive medical decisions under psychological distress. Structural constraints in genetic counseling create support gaps during waiting periods, prompting patients to rely on fragmented or misleading online information. Conversational AI may offer low-threshold, continuous informational support, yet methods for its responsible integration into clinical care remain unclear.
This study identifies and systematizes stakeholder-informed sociotechnical design requirements and governance conditions for a clinically bounded conversational AI system intended to support comprehension, appraisal, and appropriate use of complex hereditary cancer information in BRCA1/2-related care.
We conducted an exploratory, qualitative, predevelopment requirements study using 2 structured participatory workshops with interdisciplinary stakeholders (N=18). Guided exercises included stakeholder analysis, the value proposition canvas, and the sustainability awareness framework. Structured workshop outputs, field notes, and artifacts were analyzed using inductive thematic analysis.
Stakeholder workshops revealed that responsible design was shaped less by consensus around standalone responsible AI principles than by three recurrent cross-stakeholder tensions: (1) accessibility and continuity of support versus clinical role boundaries, (2) simplification and emotional reassurance versus medical precision and epistemic transparency, and (3) technical scalability versus governance, data protection, and institutional accountability. These tensions translated into design requirements for constrained system authority, mandatory human oversight, context-sensitive communication, and integration into clinical and regulatory governance structures.
Conversational AI in this context should be explicitly constrained and embedded within clinical governance structures to be perceived as responsible. Responsible implementation requires transparent design, human oversight, integration into existing care pathways, and early alignment with German/European Union (EU) regulatory requirements concerning data protection, medical device classification, genetic data, and AI transparency.
This study identifies and systematizes stakeholder-informed sociotechnical design requirements and governance conditions for a clinically bounded conversational AI system intended to support comprehension, appraisal, and appropriate use of complex hereditary cancer information in BRCA1/2-related care.
We conducted an exploratory, qualitative, predevelopment requirements study using 2 structured participatory workshops with interdisciplinary stakeholders (N=18). Guided exercises included stakeholder analysis, the value proposition canvas, and the sustainability awareness framework. Structured workshop outputs, field notes, and artifacts were analyzed using inductive thematic analysis.
Stakeholder workshops revealed that responsible design was shaped less by consensus around standalone responsible AI principles than by three recurrent cross-stakeholder tensions: (1) accessibility and continuity of support versus clinical role boundaries, (2) simplification and emotional reassurance versus medical precision and epistemic transparency, and (3) technical scalability versus governance, data protection, and institutional accountability. These tensions translated into design requirements for constrained system authority, mandatory human oversight, context-sensitive communication, and integration into clinical and regulatory governance structures.
Conversational AI in this context should be explicitly constrained and embedded within clinical governance structures to be perceived as responsible. Responsible implementation requires transparent design, human oversight, integration into existing care pathways, and early alignment with German/European Union (EU) regulatory requirements concerning data protection, medical device classification, genetic data, and AI transparency.
Authors
Lammert Lammert, Lammert Lammert, Hofenbitzer Hofenbitzer, Radtke Radtke, Sun Sun, Haimerl Haimerl, Betz Betz, Scholl Scholl, Pfeffer Pfeffer
View on Pubmed