Large language models – the artificial intelligence systems that drive tools like ChatGPT – threaten to transform health care faster than we can keep up. It’s easily the biggest tech development in medicine since the advent of CRISPr and mRNA technology.
But whether this new technology closes existing global health gaps – or stretches them further – won’t depend on capital investment or even raw technical firepower. Rather, it will depend on equity, governance, and local capacity. At least that’s the conclusion of a new analysis this week in Nature Health.
In the paper, an international group of researchers looked at how large language models are reshaping global health. They also warn that these tools could leave low- and middle-income countries behind as adoption broadens.
LLMs have already moved well beyond simple text generation. In high-income countries, they’ve become an increasingly integral part of clinical documentation, medical education, and biomedical research. And, in some cases, clinicians are relying on this technology for real-time decision support.
But the report’s authors insist that global health applications lag far behind. They point to gaps in infrastructure, a lack of local data, weak or nonexistent regulatory frameworks, and the underrepresentation of local languages and cultural contexts as huge barriers.
“While high-income countries dominate the development and deployment of LLMs, low-income countries account for less than 1% of global generative AI usage,” the authors note.
It’s a disparity that parallels longstanding digital divides in healthcare access.
Hope in Challenging Settings
Despite these challenges, the paper highlights a growing number of real-world examples where LLM-enabled systems are already improving care.
In South Africa, for example, the national MomConnect program – which has reached more than 5 million pregnant women since 2014 – is relying on LLMs to help triage health questions over text messages or smartphone apps. The system helps flag urgent cases and eases the burden on understaffed call centers.
Elsewhere, researchers have adapted transformer-based AI models for smartphone-based malaria detection, offering a scalable alternative to laboratory microscopy.
In China, a hybrid system combining image-based deep learning with an LLM has ovehauled diabetes management by supporting primary care physicians with personalized treatment recommendations.
These examples suggest that LLMs could act as “workforce multipliers,” extending the reach of overstretched health systems and improving access to care. But, the authors add, it will only work if they’re designed from a local perspective.
Bias, Hallucinations, and Hidden Costs
But the risks remain. Models trained primarily on data from high-income countries often perform poorly when applied elsewhere, introducing the risk of bias. Gaps in linguistics are especially stark. Even though the globe is home to roughly 7,000 languages, most commercial LLMs perform best in English, excluding speakers of less-represented languages.
The authors also warn about hallucinations, when these LLMs generate confident but incorrect medical information. In settings with limited clinical oversight, low health literacy, or a weak regulatory enforcement, such errors could have serious consequences. Open-weight models, which allow local deployment but often lack built-in safety guardrails, present both an opportunity and a challenge.
One other thing? The growing environmental footprint of LLMs casts a shadow over all of this progress. Training and running large models consumes massive amounts of electricity, water, and hardware. While the impact of individual queries might seem inconsequential, cumulative use across healthcare systems can rival the emissions of hundreds of households.
Open Models. Local Control.
One potential path forward lies in open-weight and open-source models, which can be deployed locally without sending sensitive health data to external servers. Models such as DeepSeek, developed under hardware constraints, have demonstrated performance comparable to leading proprietary systems at a fraction of the computational cost.
More than 300 hospitals in China have already integrated DeepSeek into clinical and administrative workflows, though the authors caution that rigorous trials evaluating patient outcomes remain scarce.
Still, the efficiency gains could be transformative for LMICs, particularly in drug discovery and research on neglected diseases, where traditional AI tools have been prohibitively expensive.
More Evidence. Better Governance.
A central concern throughout the review is the glaring dearth of hard evidence. Most evaluations of medical LLMs rely on simulations rather than randomized clinical trials. The authors call for more pragmatic trials, cost-effectiveness studies, and systematic documentation of implementation challenges.
They also argue for better communication – and collaboration – globally. Without universal safety standards, equitable data practices, and a rapid acceleration of regulatory capacity, LLM deployment risks reinforcing existing inequities rather than eliminating them. International initiatives led by the World Health Organization and others have started to address these gaps, but enforcement remains spotty.
Simply put, LLMs offer a rare opportunity to bridge divides in global health. The authors contend that they can boost frontline care, support overburdened clinicians, all while driving scalable innovation. But that potential remains unrealized.
“Without contextual sensitivity, responsible oversight, and genuine codevelopment with local partners,” they warn, “LLMs may entrench the very inequities they are meant to solve.”
The future of AI – at least in terms of global health – won’t be shaped by simple algorithms. It will be dictated by its architects, the regulators and those who reap the benefits.
Further Reading
AI Is Quietly Transforming Nursing.
Teens Are Turning to AI for Support. A New Report Says It’s Not Safe.
