Research warns: AI could widen health gap
A new critical review argues that artificial intelligence without local testing could chiefly benefit wealthy healthcare systems.
Artificial intelligence could improve access to healthcare, but also widen the health gap between wealthy and poorer countries. That is the conclusion of researchers in a new critical review, which warns about algorithms developed and tested outside local healthcare practice.
The review, published as a journal pre-proof in Health Policy and Technology, focuses on low- and middle-income countries. The authors examined existing literature on artificial intelligence in global healthcare and identified six structural risks, including inadequate local validation, biased datasets and insufficient digital infrastructure.
An algorithm can perform well in a laboratory and still be less useful in another healthcare setting. Equipment, image quality, language, patient groups and the organisation of a clinic may differ from the conditions in which the system was developed. According to the authors, such differences are still too often underestimated.
At the same time, AI can certainly have useful applications. The review mentions, among other things, support for primary care, tuberculosis screening and decision-making assistance. In areas with few doctors or diagnostic devices, a well-tested application can support certain tasks, as long as healthcare professionals provide oversight and patients are not left to an untested system.
The core of the warning is therefore not that AI is inherently harmful. The authors distinguish between a model in which AI tries to replace missing healthcare and one in which AI strengthens local healthcare professionals. The former could lead to a divide in which wealthy patients receive support from specialised systems, while poorer patients mainly receive automated answers.
A second open-access publication in npj Digital Medicine emphasises similar conditions: local control, data protection and regional networks for responsible implementation. These are policy conditions, not evidence that every existing AI system actually increases inequality.
The review is not a clinical study and provides no basis for individual medical decisions. It is a critical narrative analysis of existing research, not a systematic assessment of the effects of one specific tool. The authors outline how development is likely to unfold, in their view, if governments, funders and healthcare institutions take no countermeasures.
According to the researchers, those countermeasures consist of local validation, fair agreements on data and authorship, oversight and sustainable infrastructure. For Dutch healthcare organisations, the question is particularly relevant when they purchase AI systems trained on other populations. Transparency about performance, limitations and human control remains essential.
One story, several perspectives
What is established
- The main publication is a critical narrative review.
- The authors describe risks concerning validation, bias, data and infrastructure.
- The review also mentions potential benefits of AI in healthcare systems.
- A supplementary publication calls for local governance and regional networks.
Left
Arguments Emphasises that AI healthcare must be publicly accountable and locally controlled, with investment in staff and infrastructure rather than the replacement of human care.
Values Equal access, public control and protection against discrimination.
Consequences Fears that commercial systems will reinforce existing inequality and that poorer patients will receive inferior automation.
Centre
Arguments Wants applications to be assessed by healthcare task, with transparent validation, oversight and room for controlled innovation.
Values Evidence-based policy, proportionality and patient safety.
Consequences Emphasises that some applications could be useful, but only when their performance can be demonstrated locally.
Right
Arguments Places the emphasis on innovation, healthcare providers’ freedom of choice and removing rules that unnecessarily delay safe experimentation.
Values Technological progress, entrepreneurship and efficiency.
Consequences Fears that stringent approval requirements and public oversight will delay access to affordable diagnostics and new forms of healthcare.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The review and the supplementary publication confirm the risks and conditions described. The text clearly states that this is a critical narrative review, not evidence for individual treatment or a universal effect.
- confirmed The main publication is a critical narrative review. — The publication describes itself as a critical narrative review and journal pre-proof. source
- confirmed The authors identify six structural problems. — The abstract and highlights mention, among other things, validation, bias, data, infrastructure and division. source
- confirmed AI could have useful applications in areas including tuberculosis screening and decision-making. — The review describes these applications as existing possibilities in the literature. source
- confirmed Regional networks, local governance and data sovereignty are mentioned as conditions. — This is stated in the open-access publication in npj Digital Medicine. source
Editor's note
The main source is a critical narrative review available as a journal pre-proof; it is not a systematic review or clinical study. The conclusions concern structural risks and policy choices, not an individual medical tool.Sources
- Bridge or Chasm? The Default Trajectory of Artificial Intelligence in Low- and Middle-Income Country Health Systems — Health Policy and Technology via ScienceDirect
- Why responsible AI needs regional networks in low-resource health systems — npj Digital Medicine
More on this in Dutch media
- De Telegraaf — „kunstmatige intelligentie”
- AD — „kunstmatige intelligentie”
- de Volkskrant — „kunstmatige intelligentie”