Large language medicine needs a foundation, not only a frontier

At the Seventy-ninth World Health Assembly this year, more than 150 official statements on data, digital health, and artificial intelligence (AI) were delivered by 74 countries and 10 regional blocs, alongside 42 side events, and over 30 Ministers of Health affirming digital transformation as a core feature of resilient health systems. The through line was consistent across every region: countries are calling for concerted global leadership towards shared standards and clear governance to validate, benchmark, and regulate AI in health.Sustainable, effective AI requires an ecosystem to function, and often carries real risks to data sovereignty, privacy, and equity. AI trained on incomplete data, tested in narrow settings, or deployed without attention to bias and local context can reproduce, and even amplify, the very inequities it aims to close. It is against this backdrop that we read the lead article by Guni et al. on "large language medicine (1)." We share the authors' optimism but sharpen their central question: not whether generative and multimodal AI will reshape medicine, but how we build the scaffolding that ensures the reshaping happens safely, equitably, and sustainably.Deep learning, generative AI, large language models (LLMs), and connected agentic systems promise both incremental and transformative change across diagnostics, disease management, surveillance, prevention, and drug discovery. Yet, as Guni et al. candidly note, today's clinically mature, regulator-approved tools remain predominantly diagnostic and built on older machine learning architectures. LLMs have not yet crossed into routine practice; interconnected agentic systems are being introduced only in a fraction of cutting-edge facilities. Health systems and clinical medicine, although a rich proving ground for technical experimentation and innovation, are also understandably cautious and risk averse. The principles of evidence-based medicine further solidify the need for requisite, repeated, and rigorous demonstration of efficacy before well-established practices are altered. Some Author 1 et al. Short title/running header 4 critique this caution as innovation recalcitrance; given the duty of care in health, it is better understood as the scientific scrutiny that any clinical tool, however heavily promoted, should be expected to survive.An AI solution does not scale on its own; it scales when an ecosystem for scale exists or is built around it. This includes clear regulatory capacity and governance, strong health information systems, and trained health personnel. We posit that a key element of that enabling foundation is digital public infrastructure for health. Some of the most durable gains come not from what makes headlines but from the unglamorous parts of the system no one photographs: the unique identifiers and information exchanges, the semantic standards which allow systems to interoperate, powering supply chains that keep essential medicines on the shelf, and persistent records that follow a patient across facilities within a health system to maintain continuity of care. Countries that adopted monolithic, donor-funded digital health tools have already seen those systems collapse when funding shifted; language models deployed as parallel silos, disconnected from a shared reference architecture, risk repeating that failure at greater speed and scale. Technology-neutral reference architectures, developed collaboratively across health and telecommunications agencies and built on open standards, offer a map for where these foundations should sit (2). AI cannot pervade "every aspect of medicine," as Guni et al. envision, without an underlying layer of high-quality data structured for interoperability, disciplined evaluation protocols, and the systems analysis required to introduce AI into existing clinical workflows. Experimental systems may be introduced into clinical practice, but without clear proof points and performance benchmarks, their transition from pilot to clinical standard will likely be slow and inequitable. Ministries and hospital systems are already experiencing a glut of vendors promising impressive solutions to long-standing challenges, often without the accompanying (3,4), or countries with the least regulatory capacity will simply import externally reviewed systems without local ownership. Regulatory sandboxes, such as the United Kingdom's AI Airlock, which allows supervised testing of AI tools that fit no existing device category (5), remain an underused bridge between principle and practice. Harmonization does not mean uniformity; it means enough common evidentiary ground that a safety finding in one country still holds in another, since the models themselves will cross borders faster than any single regulator's jurisdiction can follow. Language, disease burden, health system maturity, and data availability vary enormously; a model validated in one setting cannot be assumed safe in another. Multi-institutional, privacy-preserving evaluation designed into deployment from the start, not bolted on afterward, should be the norm, particularly for populations already under-represented in training data. This brings us to what authentic inclusion in AI design actually looks like. It is not access to a shared global model. It is each country's ability to shape a model on the data of its own population, keep it under its own governance, regulate it according to its own values and context, and operate it through a health workforce with the capacity to evaluate its outputs. A ministry that cannot inspect a model, audit its performance against its own population's outcomes, or withdraw it from clinical use does not exercise meaningful oversight.Sovereignty in this context is not isolation; it is knowing where data sit, who processes them, under whose law, and on what terms access can be revoked. Without populationrepresentative data, locally grounded regulation, and a trained workforce able to question and override the system where necessary, the populations furthest behind will bear the greatest exposure to model failure, precisely because they are least represented in training data and least resourced to identify when a system is underperforming. This is a critical moment to set a shared direction for how AI in health is governed. The including an AI "commons" for shared health data assets, a verifiable AI framework grounded in clinical evidence, and the newly launched Open Health Stack Software Foundation (7), so that they remain public goods rather than proprietary platforms. The UN's Independent International Scientific Panel on AI, in its first preliminary report this year, named health among the domains where AI's benefits remain unevenly distributed and warned that safeguards cannot keep pace with its growth (8). Its wider finding, that the digitalAuthor contribution [Frontiers uses CRediT to reflect author contribution. More information is available here: https://credit.niso.org/ ] Each author must be listed under at least one of the Writing rows to qualify for authorship.Writing