AI for Health: Where can frontier AI make the biggest difference in global health?

Mapping the landscape and identifying high-impact opportunities

Research Authors: Ruby Emerson, Jenny Kudymowa, Aisling Leow, and John Firth

Summary Author: Thais Jacomassi

Artificial intelligence has become a fixture in global health conversations, but how can AI applications be leveraged to actually offer meaningful health outcomes and impact the lives of countless individuals? And from a donor’s perspective: where is frontier AI being deployed in ways that could be transformative in the healthcare sector, rather than just incremental?

In late 2025, we conducted a rapid landscape review to address these questions. We began by analyzing 258 organizations currently using AI interventions to improve clinical support, patient support, health operations, and population health. After identifying six key impact pathways through which AI could generate health value, we developed cost-effectiveness models designed to distinguish where advanced AI capabilities have the most potential. The full report¹ is accessible here².

Why this research matters

The AI-for-health landscape is vast, uneven, and moving fast. Although many organizations claim to be “AI-enabled,” not all use cutting-edge technologies to the same extent. In some cases, the “AI” in question is machine learning capabilities that have existed for decades. Genuine frontier applications, such as large language models, generative AI, and multimodal foundation models that can process text, images, and other data types together, tend to concentrate in wealthy countries and typically lack robust evidence, as these technologies are still in early deployment.

For donors and foundations working in global health, this creates a challenge: How can we identify where frontier AI can generate exceptional health benefits?

Our research provides a framework for answering this question, with concrete findings about which pathways appear most promising and which face structural barriers to impact.

What we did

1. Built a longlist of 258 organizations using AI for clinical support, patient support, health operations, and population health, using systematic searches and expert interviews.

2. Assessed the likelihood of frontier AI use for each organization based on websites, public communications, and interviews. We distinguished between traditional machine learning (standard since the 2000s), modern deep learning (mainstream since 2012), and frontier systems (LLMs and multimodal foundation models emerging since 2017). Based on these distinctions, just under half of the organizations we surveyed appeared to be using frontier AI models as of late 2025.

3. Developed cost-effectiveness models for six impact pathways:

   • Disease surveillance

   • Diagnostic assistance

   • Service delivery efficiency

   • Patient behavior

   • Clinical skills and decision support

   • Product safety and quality

4. Created illustrative profiles of five organizations that exemplified how frontier AI approaches were applied in practice.

Key findings

The frontier AI landscape

• Some “AI-powered” health organizations don’t use frontier AI. We identified 94 organizations that advertised AI capabilities but appeared to only use basic machine learning tools.

• Frontier AI is concentrated in wealthy countries, doing diagnostics and clinical streamlining. Only about one-third of the frontier AI organizations we identified operate in low- and middle-income countries (LMICs), whereas the majority focus on high-income markets. Diagnostic assistance and clinical decision support showed the highest levels of frontier AI adoption, while patient support and population health tools often relied on conventional machine learning.

• Applications of frontier AI often lack robust evaluations and data necessary for rigorous cost-effectiveness analysis.

Cost-effectiveness insights

• LMICs offered greater potential for cost-effective impact. Interventions focused on LMICs benefit from higher disease burdens, larger access gaps, and lower treatment costs.

• Throughput matters more than accuracy. Across multiple pathways, increases in service volume drove more impact than improvements in diagnostic precision. Diagnostic tools generated value by enabling more access to screening and diagnostics, rather than by marginally improving detection rates.

• Tools to improve service efficiency need to target the binding constraints. An AI tool aimed at increasing service efficiency can only increase access to care if healthcare systems are genuinely constrained by clinician time rather than budgets, infrastructure, or demand.

Most promising impact pathways, by cost-effectiveness potential

We summarize our findings regarding the five most promising impact pathways, by cost-effectiveness potential, in the table below:

Major limitations and uncertainties

• Frontier AI tools often lack strong evidence. The most advanced AI applications tend to be early-stage with limited outcome data, while better-studied interventions use older, more established AI methods that have had time to build an evidence base.

• AI is often bundled with other components of broader interventions. In many real-world scenarios, most interventions package AI with new hardware, workflow changes, training, and expanded service delivery models. Available evidence rarely isolated the contributions of AI alone.

What this means for donors

Where philanthropic support is most needed

• Disease surveillance and frontline decision support face weak commercial incentives and are more likely to require philanthropic support. In contrast, diagnostic AI markets appear to be comparatively well funded by commercial players.

• LMIC deployment requires support. Even when frontier AI tools exist, adapting them for low-resource settings and building implementation capacity often requires philanthropic backing.

Key decision factors

When evaluating AI health opportunities, we suggest focusing on:

1. Geographic focus: LMIC deployments generally offer stronger returns

2. Mechanism of action: Tools that expand access or improve throughput tend to outperform those focused only on accuracy improvements

3. System readiness: Impact depends on whether downstream systems can act on improved detection or decisions

4. Commercial viability: Support interventions with weak market incentives but strong public health value

Get in touch

For questions about this research, contact Ruby Emerson (ruby@rethinkpriorities.org) or Jenny Kudymowa (jenny@rethinkpriorities.org).

Acknowledgements

Thanks to Thais Jacomassi for writing the research summary and to Elisa Autric, Ruby Emerson, and Jenny Kudymowa for review.

Thank you!

Thank you for taking the time to read our Substack. If you would like to support our efforts, please subscribe below or share our posts with friends and colleagues.

We’re also always looking for feedback on our work. You can share your thoughts about this publication anonymously or simply reply to this email/post with suggestions for improvement or any questions.

---

By default, we’re sharing this Substack via email with Rethink Priorities newsletter members. Please feel free to unsubscribe from this Substack if you’d prefer to stick with our monthly, general newsletter.

1. Given the exploratory nature of this work and the sensitivity of some assumptions, the organization-level model outputs on cost-effectiveness are not included in the report.

2. This report was commissioned by Coefficient Giving.