AI is increasingly embedded in the institutions and environments that shape health. Yet current frameworks for understanding its implications for health equity remain underdeveloped. The social determinants of health tradition provides a strong foundation, and recent work on digital determinants of health has begun to address the health implications of digital transformation. However, AI warrants distinct conceptual attention because of its triple role: it operates simultaneously as a determinant of health in its own right, as a mediator and moderator of existing determinants, and as an amplifier of advantage and disadvantage over time. This viewpoint proposes a redistribution-translation-accumulation framework for analyzing how AI may contribute to the reproduction of health inequity. The framework comprises 2 analytically distinct mechanisms and 1 cross-cutting temporal dynamic. Redistribution captures how AI reshapes the distribution of health-relevant resources and opportunities, including education, employment, and income, while AI itself becomes an unequally distributed determinant. Translation describes how AI changes the pathways through which social positions are converted into health outcomes. Proxy-based decision rules can formalize historical inequities, diagnostic algorithms may perform unevenly across populations due to unrepresentative training data, and AI-mediated information environments can alter institutional responsiveness. Accumulation is conceptualized not as a third parallel mechanism but as a temporal amplifier operating on both mechanisms: AI-driven feedback loops and institutional embedding can concentrate advantage and disadvantage over time, often without users’ awareness. The framework is offered as a hypothesis-generating heuristic and is directionally neutral: under specifiable design, deployment, and governance conditions, the same mechanisms can narrow rather than widen health gaps in high-income and low- and middle-income settings alike. The framework has direct implications for governance. Current approaches such as the EU AI Act’s Fundamental Rights Impact Assessment (FRIA) and Canada’s Algorithmic Impact Assessment (AIA) advance AI accountability but assess systems largely before or at deployment and do not systematically track distributional health consequences. Building on this framework, I propose a distributional impact assessment as a complementary tool for equity-oriented AI governance. Structured around the 3 RTA dimensions, it asks whether an AI system alters the distribution of health-relevant resources across groups (redistribution), changes how social positions are converted into health (translation), and risks concentrating disadvantage over time through feedback and institutional embedding (accumulation). It is operationalized with candidate indicators, data sources, responsible actors, and reassessment triggers and is illustrated through a retrospective worked example of a biased care-management algorithm.

Background: As the population ages, older adults face an increasing risk of physical inactivity and related health complications, highlighting the need for scalable interventions. Smartphone-based programs have emerged as a promising strategy to support sustained physical activity among older adults. Objective: This study aimed to evaluate whether a smartphone lecture program incorporating a digital peer support app would increase physical activity among older adults, compared to a conventional smartphone lecture program. Methods: This 2-arm, 1:1 parallel-arm, cluster-randomized trial was conducted in 2 urban regions of Japan (Sumida Ward, Tokyo, and Chiba City, Chiba). Eligible participants were community-dwelling adults aged ≥60 years, able to walk independently, and smartphone users; exclusion criteria included prior use of the peer support app or medical restrictions on walking. Participants were recruited offline during community smartphone lectures (closed-group recruitment). The intervention combined face-to-face lectures with app-based peer support, while outcomes were assessed both objectively (via smartphones) and through self-administered paper questionnaires. All participants received a baseline smartphone lecture. Intervention participants attended 2 additional sessions using a digital peer support app (Minchalle; A10 Lab Inc), which included features such as daily step goals, peer sharing, and group encouragement. Control participants attended 2 standard follow-up smartphone lectures. The primary outcome was the change in weekly average daily step count from baseline to Week 12. Secondary outcomes included total metabolic equivalent of task (MET)–minutes per week (assessed via the International Physical Activity Questionnaire), walking time (≥30 minutes per day), daily smartphone use, and number of smartphone use purposes. Results: A total of 156 community-dwelling older adults were grouped into 40 clusters and randomized (20 intervention clusters, n=80 and 20 control clusters, n=76). In total, 124 participants (79.5%) completed the follow-up, and valid step data were available for 117 participants, with missing data ranging from 5.1% to 29.1%. Baseline daily steps averaged 3951 (SD 1686) in controls versus 4583 (SD 1973) in the intervention arm. An unadjusted mixed model for repeated measures showed significantly higher step changes for intervention participants at Week 12 (difference=579, 95% CI 36-1123; =.04). No significant differences emerged for total METs (difference=646 MET-min per week, 95% CI –12 to 1303; =.054) or walking ≥30 minutes per day (odds ratio [OR] 1.56, 95% CI 0.63-3.90; =.33). However, the intervention arm demonstrated a significant increase in daily smartphone use (OR 4.10, 95% CI 1.15-14.6; =.03) and in the number of smartphone use purposes (difference=0.58, 95% CI 0.12-1.05; =.01). Conclusions: A smartphone lecture program integrated with app-based peer support led to modest but meaningful improvements in step counts among older Japanese adults, at Week 12 of the 12-week intervention. Future research should investigate long-term maintenance, additional measures of physical activity, and subpopulation responses to optimize digital health programs for older adults. Trial Registration: UMIN-CTR UMIN000051904; https://tinyurl.com/4m4zm99v