Promise to Practice: Reimagining Artificial Intelligence for Equitable Global Health Impact

A landscape in motion: Current state of AI in global health

The AI working group reviewed four global health domains, education, epidemiology/surveillance, clinical medicine, and data equity, to identify major applications, stakeholders, implementation challenges, evidence gaps, and illustrative case examples. Findings revealed a wide range of AI applications, from personalized education to predictive analytics, and highlighted the fragmented nature of deployment [8–10]. Despite domain‑specific differences, several common themes emerged across settings (Table 1).

Education: AI applications in global health education include adaptive learning platforms, automated feedback systems, simulation technologies, and curriculum analytics [11]. Early evidence suggests improvements in engagement and personalization, yet robust longitudinal evaluations linking AI‑supported education to workforce performance or patient outcomes remain scarce [12–15]. Most published evaluations remain short‑term or descriptive, leaving uncertainty about effects on workforce competency, clinical behavior, or patient outcomes. Challenges include unequal access to hardware and connectivity, limited faculty preparedness, and unresolved ethical questions around data privacy [10].

In many low‑ and middle‑income country (LMIC) settings, AI‑enhanced education risks widening disparities if infrastructure and digital literacy gaps persist. Key stakeholders include educational institutions, healthcare providers, technology developers, students, educators, policymakers, and funding agencies.

Epidemiology and surveillance: AI tools increasingly support outbreak detection, predictive modeling, and digital contact tracing. Platforms like BlueDot (Canada), HealthMap (Boston), Arogya Setu (India), and BioSense (US CDC) illustrate diverse deployment models. Although technical feasibility is often demonstrated, few longitudinal studies assess model degradation, false‑positive burden, cross‑border data governance, or comparative cost‑effectiveness relative to conventional surveillance [8, 16]. Funding primarily comes from government agencies (NIH, CDC, DARPA, EU, PHAC, UKRI) and philanthropic organizations (Gates Foundation, Wellcome Trust), which may limit the long‑term sustainability of these efforts.

Clinical medicine: AI applications include education augmentation aides, diagnostic imaging support, decision‑support systems, ambient documentation tools, drug discovery, and virtual care optimization [17]. AI‑assisted tools may support asynchronous learning and remote care workflows, but evidence of broad clinical benefit remains uneven across settings [11]. Additionally, ambient scribes and related tools aimed at reducing administrative and cognitive burden are being adopted with increased frequency in various clinical settings [17]. Technical accuracy often precedes implementation evidence, and generalizability remains a significant concern. Implementation challenges, including provider trust, workflow integration, and accountability frameworks, often dominate impact. Reported benefits include improved diagnostic performance in selected use cases, more efficient resource allocation, and accelerated analytic workflows; however, these gains do not consistently translate into real‑world effectiveness across settings. Several highly accurate models also fail at the point of implementation because workflow disruption, poor calibration in new populations, unclear liability, or dependence on infrastructure is unavailable in low‑resource environments.

Data equity: Across domains, concerns regarding algorithmic bias, dataset imbalance, and limited transparency persist [8]. Here, data equity includes who is represented in datasets, who governs access to those data, and who benefits from the resulting tools. AI systems may reproduce structural inequities embedded within training data and governance architectures [18, 19]. Although participatory design methods and fairness auditing frameworks are increasingly proposed, regulatory enforcement and accountability mechanisms remain inconsistent, particularly across LMIC contexts [20]. In practice, participation may remain consultative rather than power‑sharing, limiting the corrective potential of these approaches.

Conference session insights

To share the findings of the landscape review and foster dialogue on AI in global health research, a pre‑conference session, “Harnessing Technology and AI for Collaborative Breakthroughs in Global Health,” was held at the 2025 CUGH conference. The session opened with four “Setting the Scene” presentations: the AI Working Group’s landscape review, results from a CUGH Trainee Advisory Committee survey of emerging global health professionals, a review on predictive AI applications in African settings, and a review of AI policies in LMICs.

Across the presentations, a consistent theme was uneven preparedness for AI adoption in global health. Trainees reported interest in formal AI education but limited confidence in current use, while examples from low‑resource settings showed that implementation barriers, governance gaps, and infrastructure constraints often outweighed technical promise. These observations reinforced the need for workforce development, context‑sensitive tool selection, cross‑industry education resource development, and a stronger policy and governance framework to mitigate the potential of AI to reinforce, perpetuate, and embed inequities.

Innovation in practice: Case studies across contexts

The session’s second part featured five case studies that moved the discussion from principle to practice (Table 2). For example, a study from Uganda showed that imaging‑based diagnostic support may perform well analytically yet remain constrained by electricity reliability, device maintenance, connectivity, and local workflow integration. Similarly, AI‑enabled cancer screening across multiple LMIC settings highlighted the dependence of image‑classification tools on large, well‑annotated, and locally representative datasets. Additionally, AI‑powered systems illustrated potential in improving healthcare service delivery and clinical decision support. Across cases, the central lesson was that contextual readiness, governance, and implementation feasibility often determine impact more than model performance alone.

Key lessons emerged:

• Context is critical. Tools developed in high‑resource environments often assume stable internet, interoperable data systems, maintenance capacity, and price points that do not reflect many low‑resourced contexts.

• Success is not just technical. Implementation challenges, not model performance [21], were often the primary barriers to impact.

• Ethical concerns persist. From iris scanning technologies to AI‑assisted triage, discussions revealed ongoing tensions around consent, privacy, surveillance, and accountability, especially in humanitarian, fragile, or other vulnerable settings.

These examples underscored a paradox: the greatest promise of AI lies in low‑resource settings, yet these are also the places most structurally excluded from designing, implementing, and benefiting from it [16, 18, 22, 23]. However, recent advances in edge technology and “low‑tech” native solutions are increasingly designed and utilized by low‑resourced settings [24].

Expert dialogue: Synthesis and future directions

The session ended with a moderated discussion that reinforced three cross‑cutting themes: AI in global health should be co‑led by public health and domain experts rather than framed as a purely technical agenda; implementation evidence matters as much as algorithmic performance; and governance, digital literacy, and infrastructure remain decisive constraints on equitable deployment. Central to this is the urgent need for enhanced AI and digital literacy across the entire health ecosystem and strong ethical governance standards.

Participants converged around five key core priorities for responsible AI deployment:

1. Establishing robust governance structures and local policymaking to protect countries from exploitation and harm;

2. Comprehensive capacity‑building that extends to the entire health‑technology ecosystem, including not only AI literacy but also technical and regulatory training;

3. Commitment to human‑ and community‑centered and equity‑oriented design, requiring the development and implementation of AI tools with and for intended users;

4. Use of inclusive training data beyond high‑income datasets emphasizing underrepresented populations by adapting existing models or developing local ones;

5. Development of the evidence base for safe, locally responsive, and effective AI deployment in diverse global settings through rigorous, real‑world evaluation, including randomized trials and cost‑effectiveness studies (Table 3). Without this evidence base, AI risks scaling technical performance without demonstrable population‑level benefit.

Addressing these priorities requires the development of novel, collaborative governance models that integrate expertise from public health, ethics, law, and human rights to guide cross‑border AI initiatives.

These principles align with emerging global frameworks and help identify practical next steps for more equitable AI deployment in global health. Particular emphasis must be placed on LMIC leadership and co‑creation in AI deployment, especially in the high‑stakes field of health. Lessons from this workshop and the CUGH collaborative approach can be translated to other partnerships involving LMICs, particularly when deploying novel AI technologies.

The session also underscored the importance of coalition building, LMIC leadership, and practical guidance for implementation. Diverse participation strengthened the relevance of the discussion, while barriers such as travel costs and visa constraints highlighted how inequities in convenings can shape whose knowledge is represented. These observations reinforce the need for inclusive governance, context‑specific implementation tools, hybrid‑by‑design convenings, and stronger mechanisms for shared decision‑making.

A call to action: Five strategic imperatives

Based on the session insights and comprehensive literature review, we propose five strategic imperatives for advancing AI in global health (Table 4):

1. Develop and disseminate contextualized AI governance frameworks

   Collaborate with WHO, regional bodies, and LMIC governments to establish adaptable, enforceable AI guidelines that address ethics, consent, data ownership, accountability, and local implementation realities [4, 6, 7, 20, 25]. Our review revealed significant regulatory gaps, with LMICs particularly vulnerable to untested implementations. Contextualized governance must be tailored to local regulatory, infrastructural, and sociopolitical realities.

2. Invest in equitable capacity‑building

   Support digital infrastructure, local data ecosystems, and AI training initiatives that empower local actors to build, adapt, and evaluate tools. Address the digital divide that currently limits access to AI benefits in under‑resourced regions, and integrate AI literacy into health professional education [10, 26, 27].

3. Fund implementation research and evidence generation

   Ensure funding bodies prioritize rigorous evaluation and real‑world deployment studies over purely technical performance metrics. Priority areas include pragmatic trials, cost‑effectiveness, post‑deployment monitoring, and comparative validation across diverse contexts.

4. Create open‑access knowledge repositories

   Share curricula, code, datasets, and toolkits under open science principles where appropriate to prevent duplication and promote transparency, while guarding against data exploitation and respecting data sovereignty and intellectual property rights.

5. Center community engagement, ethics, and equity

   Design AI systems in dialogue with the communities they serve. Value indigenous knowledge, uphold data sovereignty, ensure benefits are reciprocal, and involve affected communities in priority setting, design, implementation, and oversight [2, 8].