JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) is increasingly transforming health communication by enabling scalable, multilingual, and personalised information delivery. While AI systems are often promoted as tools to improve access for culturally and linguistically diverse populations, emerging evidence suggests that their performance varies significantly across languages and contexts. In particular, concerns have been raised about a “cultural alignment deficit,” whereby AI systems achieve technical accuracy but fail to account for cultural meaning, communication norms, and contextual relevance, potentially reinforcing existing health inequities.

Objective:

This review aims to examine the current state of practice in linguistic and cultural adaptation in AI-enabled health communication, and to evaluate the effectiveness of these systems across different linguistic and cultural contexts.

Methods:

A scoping review was conducted following PRISMA-ScR guidelines. Four major databases (PsycINFO, Scopus, Web of Science, and PubMed) were searched, supplemented by Google Scholar and grey literature sources. Studies were included if they examined AI-based health communication systems with a focus on linguistic or cultural adaptation. Both empirical studies and review-type papers were included. Data were extracted using a standardised form, and findings were synthesised narratively across key dimensions, including adaptation approaches, system performance, and patient-centred outcomes.

Results:

A total of 104 studies were included, comprising 52 empirical studies and 52 review or conceptual papers. The findings reveal a strong Anglocentric bias, with most research conducted in English-speaking or Western contexts. Linguistic and cultural adaptation was frequently limited to surface-level translation, with minimal integration of cultural norms, metaphors, or contextual meaning. AI systems demonstrated improvements in readability, efficiency, and patient engagement, particularly in high-resource languages such as English and Spanish. However, performance declined in low-resource and non-English contexts, where outputs were more likely to contain inaccuracies, reduced cultural resonance, and lower perceived trustworthiness. Hybrid human–AI approaches, involving clinician or cultural expert input, were consistently associated with improved outcomes. Overall, evidence for long-term effectiveness, trust, and equity remains limited.

Conclusions:

AI-enabled health communication shows considerable potential to improve access and efficiency, but its benefits are unevenly distributed across linguistic and cultural contexts. Current systems remain largely constrained to surface-level adaptation, failing to achieve deeper cultural alignment. This review highlights the need to reconceptualise AI not merely as a translation tool but as a culturally competent communication partner. Embedding cultural adaptation at the design stage, improving methodological transparency, and addressing structural inequities in data and system development are critical to ensuring equitable and trustworthy AI-mediated health communication.