WHO moves healthcare AI from principles to practice

The international debate about artificial intelligence in healthcare is shifting from principles towards implementation. At a three-day meeting in Hangzhou, the World Health Organization (WHO), International Telecommunication Union (ITU) and World Intellectual Property Organization (WIPO) are bringing together policymakers, regulators, healthcare professionals, researchers and technology experts to examine how AI can be introduced responsibly into health systems.

The third meeting of the Global Initiative on AI for Health (GI-AI4H), taking place from 16 to 18 September, has practical, responsible and sustainable implementation as its central theme. The programme covers governance, international cooperation and technical priorities, but also looks specifically at AI applications already deployed in healthcare. Combined with new guidance from WHO, ITU and WIPO on bringing AI-enabled health innovations from development to deployment and scale, the message is becoming clear: demonstrating what AI can do is no longer enough.

Real-world implementation becomes the test

GI-AI4H has invited organisations to submit examples of AI applications already deployed in health services or health systems. The initiative is particularly interested in applications contributing to primary healthcare, universal health coverage or health-system strengthening. Importantly, submissions are expected to describe not only the technology, but also its scale, results and supporting evidence, governance arrangements, accessibility and the problems encountered during implementation.

That approach addresses one of the persistent weaknesses in healthcare AI. Algorithms can perform impressively in research environments, yet implementation introduces an entirely different set of challenges. Clinical responsibility has to be defined, technology needs to fit existing workflows, professionals need to trust and understand it, infrastructure has to support it and organisations need evidence that benefits survive beyond a controlled pilot.

The selected cases are intended to provide practical lessons for health authorities, policymakers, researchers and organisations implementing AI. WHO, ITU and WIPO explicitly state that selection does not constitute endorsement of a particular technology, institution or approach.

The same implementation focus is visible in the organisations' recently published guide, AI-enabled Health Innovation and IP: From idea to impact. It follows AI-enabled healthcare innovation through the full lifecycle, from ideation and model training to validation, regulatory approval, deployment and international scaling. Intellectual property is part of that journey, but so are health data governance, technical standards, regulatory pathways, partnerships and commercialisation.

Testing AI without surrendering clinical data

GI-AI4H is also exploring another practical barrier through its international Benchmarking Challenge: how hospitals and other institutions can evaluate AI models using clinical data without handing those data to the model developer. At the same time, developers do not have to disclose their models to the institution providing the evaluation data.

Models are tested in their existing form rather than being trained on the participating institution's data. Healthcare organisations can contribute clinical tasks and datasets and determine appropriate performance metrics. Solutions are assessed for correctness, non-disclosure and cost, while subgroup results can be reported where the available data allow this.

The approach is relevant because healthcare organisations increasingly need independent evidence about AI performance without weakening control over sensitive patient information. It could provide a route towards evaluating models under clinically meaningful conditions while keeping data within institutional boundaries. The initiative nevertheless stresses that benchmark results do not constitute regulatory approval, certification or endorsement.

From invention to impact

The involvement of WIPO adds another dimension to the implementation debate. Its new joint publication with WHO and ITU shows how intellectual property, data, regulation and technical standards become interconnected as an AI application moves towards the market. An AI-enabled healthcare product may involve patents, software copyright, proprietary datasets, trained models, trade secrets and licensing agreements, all of which can affect partnerships and the ability to scale internationally.

The guide's Innovation Lifecycle IP Matrix maps these considerations across different development stages. Early decisions may concern ownership, documentation and confidential information; later stages introduce questions about training data, validation partners, regulatory requirements, licensing and entry into new markets. The underlying point is that scaling healthcare AI is not simply a technical exercise.

That makes the Hangzhou meeting particularly relevant. Healthcare AI is entering a phase in which model performance remains important, but increasingly sits alongside governance, evidence, data control, regulation, standards and viable implementation models.

For healthcare leaders, the next question is therefore not which AI application produces the most impressive demonstration. It is which applications can make the much harder journey from promising technology to trusted, scalable healthcare.