Background:
Culturally and linguistically diverse (CALD) patients experience inequitable access to cancer information. Patients may rely on unverified artificial intelligence (AI) tools for information. These tools can provide inaccurate information, compromise health literacy and affect care. Safe implementation of AI in health requires consumer-centred design, robust clinical governance, rigorous multilingual validation coupled with technological transparency to ensure trustworthy information.
Methods:
Using experience-based co-design, consumers from Arabic, Chinese, Serbian and Vietnamese-speaking communities, healthcare providers and organisational stakeholders participated in multiple workshops to identify information needs and barriers plus requirements for AI-enabled cancer support. Feedback directly informed the development of a multilingual chatbot using NSW Cancer Institute eviQ treatment protocols. This co-design framework allowed iterative refinement to improve empathy, usability and safety. Multilingual consumers evaluated the chatbot using simulated cancer patient scenarios. Outputs were qualitatively assessed by consumers, clinicians and NAATI-accredited translators for usability, clinical accuracy, cultural appropriateness, translation quality and safety.
Results:
Co-design workshops identified the following key themes: language-related informational inequity, cautious acceptance of AI, preference for AI as an adjunct to clinician care, unmet cultural and supportive care information needs. Post implementation of feedback, the final consumer acceptability exceeded 80% across clarity, usability, trust, safety and cultural appropriateness domains. Across translator evaluations, Serbian and Chinese chatbots consistently performed well, with greater variability in Vietnamese and Arabic translations, highlighting language-specific implementation challenges. Clinically relevant errors were uncommon (<5%), and no high-risk translation errors were identified. Clinician validation of medical accuracy and safety resulted in implementation of specific safeguards, including a provision for mental health concerns.
Conclusions:
This study demonstrates a practical framework for designing and implementing clinically governed AI tools for CALD communities. Combining co-design feedback loops and structured safety validation, we have developed a framework to achieve equitable access to trustworthy cancer information while identifying and mitigating language-specific risks.