Application and Integration of Generative AI in Medical Education Vary Across Regions: Why Is This an Important Consideration?

By Farooq Azam Rathore¹, Fareeha Farooq²

Affiliations

1. Department of Rehabilitation Medicine, Quetta Institute of Medical Sciences, Quetta, Pakistan
2. Department of Biochemistry, Quetta Institute of Medical Sciences, Quetta, Pakistan

doi: 10.29271/jcpsp.2026.08.1095

Sir,

We have read the viewpoint “An Overview of Generative Arti- ficial Intelligence (AI) in Medical Education”  by Wang et al.¹ with interest. They have discussed the technological advancements behind the recent rise of generative AI, including the evolution of large language models (LLMs) and immersive simulation platforms. Examples of innovative applications, such as virtual synthetic patients, automated case scenario generation, and advanced feedback systems, highlight the potential of GAI in contemporary medical education.

While the viewpoint is comprehensive, its focus and examples are predominantly drawn from the global North and countries with well-established, centuries-old medical education systems, abundant resources, and readily available technology. Many of the examples mentioned are not currently available or applicable to developing regions, which comprise the majority of the world's population. This limitation is important in light of the growing body of evidence from low‑ and middle‑income countries (LMICs) supporting the integration of generative AI in medical education. Studies from Pakistan,² Egypt,³ India, and Sierra Leone⁴ indicate that faculty and students generally have positive attitudes towards the use of generative AI in medical education and are willing to adapt it. However, these studies also highlight several challenges, including inadequate digital infrastructure, limited foundational training in artificial intelligence, and ethical concerns such as data privacy, security, and algorithmic  bias.

The viewpoint mentions innovations such as virtual reality anatomy platforms and custom-trained models. However, these technologies often rely on high-performance computing, stable electricity, and high-speed internet. These facilities are not widely available in many LMIC settings. In Pakistan, for example, medical colleges in small peripheral cities often operate without reliable digital infrastructure or dedicated IT departments, and students and faculty are not formally trained in AI literacy. Without addressing these constraints, there is a risk of reinforcing global disparities in access and quality of medical  education  due  to  AI.

The authors highlight the potential of open-source models such as LLaMA. However, the implementation of even light-weight LLMs remains a challenge in LMICs where access to high-performance GPUs is limited or nonexistent. Another issue is the limited discussion on contextually relevant data-sets. The training data for most LLMs are derived primarily from Western healthcare settings and may have limited relevance to the unique sociocultural, economic, and healthcare contexts of LMICs.⁵ Western-centric training data leads to biases against LMICs in healthcare-related decisions.⁶ This raises concerns about socio-cultural bias and diagnostic irrelevance when applied in LMICs. Without access to locally relevant, appropriately annotated, linguistically diverse, and ethically sourced data, AI systems risk perpetuating diagnostic inaccuracies and generating culturally inappropriate recommendations.

Despite these limitations and challenges, faculty and students in the LMICs must not be excluded or left behind in the integration of generative AI in medical education. There can be multiple ways to address this. Governments and national regulatory bodies should invest in developing AI-compatible digital infrastructure. Universities and institutes should arrange formal training to improve AI literacy in faculty and students. The faculty and students themselves must make the best use of high-quality free resources available for training in prompt engineering and the ethical use of AI. Institutions and faculty can begin with simple initiatives, such as using LLMs and AI chatbots to support lesson planning and generate outlines for lectures and assessments. Future studies should focus on developing locally relevant datasets and establishing standardised assessment metrics that reflect the diverse realities of medical education across different regions and countries. This will ensure that the deployment of GAI in medical education does not disproportionately benefit only a few insti- tutions and individuals in high-resource settings; instead, its benefits can be shared equitably across the globe.

COMPETING  INTEREST:
The  authors  declared  no  conflict  of  interest.

AUTHORS’  CONTRIBUTIONS:
FAR:  Conception  and  drafting.
FF: Literature review and revision of the draft for critical input.
Both authors approved the final version of the manuscript to be published.

REFERENCES

1. Wang S, Geng R, Xu R. An Overview of generative artificial intelligence in medical education. J Coll Physicians Surg Pak 2025; 35(6):793-6. doi: 10.29271/jcpsp.2025.06.793.
2. Naseer MA, Saeed S, Afzal A, Ali S, Malik MGR. Navigating the integration of artificial intelligence in the medical education curriculum: A mixed-methods study exploring the perspectives of medical students and faculty in Pakistan. BMC Med Educ 2025; 25(1):273. doi: 10.1186/s12909-024- 06552-2.
3. Ghanem OA, Hagag AM, Kormod ME, El-Refaay MA, Khedr AM, Abozaid OM, et al. Medical students' knowledge, attitudes, and practices toward generative artificial intelli-gence in Egypt 2024: A cross-sectional study. BMC Med Educ 2025; 25(1):790. doi: 10.1186/s12909-025-07329-x.
    
4. Choi JH, Garrod O, Atherton P, Joyce-Gibbons A, Mason-Sesay M, Bjorkegren D. Are LLMs useful in the poorest schools? the teacher AI in Sierra Leone. arXiv 2023; Available from: httpss://arxiv.org/abs/2310.02982.
5. Navigli R, Conia S, Ross B. Biases in large language models: Origins, inventory, and discussion. ACM J Data Inf Qual 2023; 15:1-21. doi: 10.1145/3597307.
6. Manvi R, Khanna S, Burke M, Lobell DB, Ermon S. Large language models are geographically biased. 2024. Available from: httpss://arxiv.org/abs/2402.02680.