AI as a saviour

The Danger of Assuming AI Works Everywhere

•

A recent World Bank report1 highlights the promises of artificial intelligence (AI) and illustrates various examples of how AI can be a “saviour”, particularly for developing contexts around the world. Here are the main takeaways from the almost 400-page report. While the report makes a compelling case for the potential of AI, I think there is a bigger question we need to ask: what happens when these tools are developed without sufficient consideration of the contexts in which they will actually be used? 

One of the most common concerns with AI is the threat of losing jobs to it – according to the World Bank, developing contexts will lose less than 5% of jobs to AI, compared to almost 15% in developed contexts. This is because a smaller share of jobs is susceptible to immediate automation. Consequently, the report notes that AI can be used to rapidly bridge acute skill shortages in various sectors, and several use cases of AI being used positively are highlighted. For instance, in India, AI weather forecasting yielded net savings up to $560 per farmer, and in Ghana, the Rori chatbot delivered a year of math learning via SMS for $5/student. 

In all honesty, I can’t deny the lucrative nature of AI tools when it also promises health equity, better access for the most vulnerable or underserved, and a reduction in workload for our already overburdened workforces. My biggest concern, however, remains that AI models are often not adapted to local contexts. If you know how AI algorithms work, then you know that LLMs (Large Language Models) are trained on a particular subset of data. In most cases, this data is based in the Global North, meaning that its solutions are applicable to a very specific type of population. In fact, a piece in the World Economic Forum estimates that 5 billion people could be systematically excluded by using AI in this manner.2 

Furthermore, LLMs have struggled to account for cultural and contextual differences. For instance, the report notes that an AI health tool trained in wealthy countries systematically over-recommended costly lab tests when deployed in Nigeria because it assumed Western medical norms. This is an example of an unfair financial burden, but there are examples of medical misdiagnoses too, such as when algorithms built to detect skin cancer or cardiovascular risk have been trained on primarily light-skinned or European and American populations, respectively; when a COVID-19 triage model trained in Oxford failed in Vietnam due to demographic differences; or when an eye-screening AI tool in Thailand rejected 20%+ of images because clinic lighting didn’t match training data.

To its credit, the report does note the importance of adopting and adapting AI models to local contexts, which is vital in order to reduce biases and systemic injustices. However, it would be more meaningful to see funding models represent this strategy.

This is not to say that there are no good implications of AI in healthcare. One of the examples presented by the report is in Bangladesh, where such technology increased screening for diabetes related eye conditions by almost 40%. Similarly, the Indian app Cough Against TB correctly diagnosed 9 out of 10 TB patients. However, my concern with this kind of a ‘saviour’ approach is that we are once again thinking of AI as a ‘be-all, end-all’ solution to challenges faced by developing contexts, when it may not actually be that. 

For instance, one of the ‘promises’ of AI in this report is that it can aid in expanding the supply of healthcare workers. In theory – yes – it can. Regardless, such claims may only be looking at healthcare workers from one perspective, that of having medical knowledge. In practice, however, we know that healthcare workers require various skills. One being theoretical medical knowledge, another being bedside manner, considering contextual or cultural factors when relaying medical information to patients, and knowledge that is learned through experience or trial and error, rather than what you can find in a textbook or academic journals. 

While I agree that AI can most definitely be a useful assistant to reduce the workload of overburdened healthcare workers, particularly in LMICs, various examples show the potential harm it can cause when AI tools are relied on blindly or without human support. For instance, a Johnson and Johnson product aimed at assisting surgeons treating chronic sinusitis is reported to have had over 100 malfunctions and adverse events since the addition of an AI assistant, of which stroke was one of them.4 

Again, this is not to alarm readers. My argument here is simply that AI models should not be used en masse without evaluating their correct uses or validating them. A study by Qiao Jin et al came to the same conclusion.5 Another consideration is that legislation does not yet have guidelines on what happens in the event of adverse events when AI tools are involved. The Medical Protection Society recently warned doctors and the NHS that they could be the target of lawsuits should such events occur.6 

Having taught undergraduate students myself in the age of AI, we must also take into account the risk of cognitive offloading, which the report also addresses. Cognitive offloading is whereby students rely on AI tools to do their thinking for them on the basis that AI tools have more knowledge than they could have. Not only does this diminish critical thinking capacities, but it also takes away true learning experiences that can strengthen students’ real-world interactions when they go on to do research or practice as healthcare workers.

The World Bank report isn’t healthcare-specific, and outlines the promises and limitations for several other industries. Importantly, it has a chapter on the role of governments in ensuring trustworthy AI tools through regulation and contextually relevant frameworks. This is particularly important when we consider that the consequences of poorly designed or poorly adapted AI tools are unlikely to be distributed equally. 

Ultimately, I don’t think the question is whether AI has a place in healthcare, education or other sectors. It clearly does, and there are already examples of it being used to improve access, reduce workloads and address gaps in healthcare and other services. The question is how we use it, who it is designed for, and what happens when it doesn’t work as intended.

If AI tools are developed using data that does not adequately represent the populations they are intended to serve, then deploying them without adapting or validating them locally risks perpetuating the same inequalities that we are trying to address. In this sense, AI does not exist outside of the systems in which it is developed and used. It can reflect existing biases in our healthcare, education and social systems, and potentially reinforce them at scale.

Perhaps this is why the idea of AI as a ‘saviour’ concerns me. The technology itself is not necessarily the problem. It is the assumption that a technological solution can be applied to complex, context-specific problems without the need for human expertise, local knowledge, adaptation, regulation or accountability. AI can be a useful tool, but it cannot be separated from the people, systems and inequalities that surround it.

References

  1. World Bank. World Development Report 2026: The Promise of Artificial Intelligence [Internet]. Washington (DC): World Bank; 2026 [cited 2026 Sep 27]. doi:10.1596/978-1-4648-2331-2 
  2. Onuh G. AI in healthcare risks could exclude 5 billion people; here’s what we can do about it. World Economic Forum [Internet]. 2025 Oct 1 [cited 2026 Sep 27]. Available from: https://www.weforum.org/stories/health-and-healthcare-systems/ai-in-healthcare-risks-could-exclude-5-billion-people-here-s-what-we-can-do-about-it/
  3. Tao Y, Viberg O, Baker RS, Kizilcec RF. Cultural bias and cultural alignment of large language models. PNAS Nexus. 2024;3(9):pgae346. doi:10.1093/pnasnexus/pgae346.
  4. Reuters. As AI enters the operating room, reports arise of botched surgeries and misidentified body parts. Reuters [Internet]. 2026 Feb 9 [cited 2026 Sep 27]. Available from: https://www.reuters.com/investigations/ai-enters-operating-room-reports-arise-botched-surgeries-misidentified-body-2026-02-09/
  5. Jin Q, Chen F, Zhou Y, Xu Z, Cheung JM, Chen R, Summers RM, Rousseau JF, Ni P, Landsman MJ, Baxter SL. Hidden flaws behind expert-level accuracy of multimodal GPT-4 vision in medicine. NPJ Digital Medicine. 2024 Jul 23;7(1):190. 
  6. Campbell D. Doctors and NHS could be sued for mistakes made by AI tools, report warns. The Guardian [Internet]. 2026 Jun 9 [cited 2026 Sep 27]. Available from: https://www.theguardian.com/society/2026/jun/09/doctors-nhs-could-be-sued-mistakes-ai-tools-medical-protection-society-report

Leave a Reply

Your email address will not be published. Required fields are marked *