Who Holds Responsibility in AI-Driven Care?

Artificial intelligence is often presented as a way to make health care faster and more accurate. Whether it also makes care more equitable will depend on how these systems are developed, tested, and distributed.

One major concern is data representation. AI models learn from existing health data, which may underrepresent certain racial, socioeconomic, linguistic, and age groups. Historical data can also reflect disparities already embedded in the health care system. If a model treats those patterns as objective, it may reproduce them in its predictions rather than correct them.

Access presents another challenge. Well-funded health systems are better positioned to purchase and test new AI tools, while hospitals serving under-resourced communities may be excluded from their benefits. This could create a system in which the patients facing the greatest barriers to care also have the least access to technologies designed to improve it.

Still, thoughtful implementation shows how AI could strengthen care. When Mass General Brigham introduced an AI tool that drafts clinical notes from patient-provider conversations, it tested the technology across different accents, languages, and clinical environments. Among the first 220 participating providers, the system was associated with a 40 percent relative reduction in burnout. Nearly 80 percent also reported being able to pay more attention to patients during visits.

For policymakers, regulating health care AI should involve more than evaluating its average accuracy. Standards should require developers and health systems to report how tools perform across demographic groups, languages, and care settings. Systems also need continuous monitoring, since AI models and their outputs can change over time. Public investment may be necessary to prevent cost from limiting access among safety-net and rural providers.

AI will not automatically reduce health disparities. It can improve care, but only if equity is treated as a requirement from the beginning rather than something assessed after deployment.

Source: Harvard Medical School.