Responsible Artificial Intelligence Integration in Medical Student Education in Somalia and Low-Resource Settings: A Context-Sensitive Implementation Framework Authors: Mohamed Mohamud Ali, Abdullahi Abdirahman Omar Published: 2026 ABSTRACT AbstractArtificial intelligence (AI), particularly generative AI and large language models, is increasingly used by health-professions learners for explanation, revision, simulation, feedback, and assessment support, while institutional governance has not always kept pace. Somalia is a distinctive setting in which recent mapping identified 112 health-professions schools, many concentrated in urban areas, while medical education continues to face unequal access to faculty, simulation, standardized assessment, and reliable digital infrastructure. This commentary proposes a context-sensitive implementation framework for responsible AI integration in medical student education in Somalia and comparable low-resource settings. The framework was developed through a focused narrative synthesis of Somalia-specific educational evidence, peer-reviewed AI-in-medical-education literature, and international guidance from WHO, UNESCO, and the Association of American Medical Colleges. Its contribution is to translate broad AI principles into a resource-constrained implementation pathway that links Somalia-specific educational constraints to defined AI-supported functions, governance safeguards, phased institutional actions, and measurable educational outcomes. Published studies support the feasibility of AI-assisted feedback, virtual patient simulation, and tutoring, but also demonstrate variable accuracy and the need for human verification. The proposed model therefore prioritizes faculty oversight, academic integrity, patient confidentiality, multilingual verification, local clinical validation, low-bandwidth access, and outcome monitoring. A phased roadmap moves from policy and low-risk pilots to curriculum integration and multi-institutional evaluation. The framework is conceptual and requires prospective validation. If implemented with simultaneous investment in faculty development and assessment reform, AI may expand access to supervised practice without replacing teachers, patients, bedside learning, or professional judgment. Introduction Medical education in low-resource settings must produce clinically competent, ethical, and socially accountable physicians despite constraints in faculty availability, simulation capacity, clinical exposure, assessment infrastructure, and digital access. These pressures are particularly relevant in Somalia, where decades of conflict and health-system fragility have coincided with rapid growth in health-professions education. A recent review identified 112 health-professions schools across Somalia and documented marked urban concentration, variable regulation and accreditation, workforce shortages, and maldistribution of skilled health professionals.1 Artificial intelligence (AI) is an umbrella term for computational systems that perform tasks associated with human intelligence. Generative AI refers to models that create new content, while large language models (LLMs) are a class of generative AI trained to generate and interpret language. Reviews of AI in medical education describe applications in individualized learning, assessment, simulation, feedback, curriculum support, and learner engagement, but also emphasize uncertain reliability, bias, privacy, academic-integrity risks, and uneven evaluation.2–4 Most published AI-in-medical-education literature has originated from higher-resource environments, where institutional infrastructure and specialist faculty are more available.2 The policy question is already immediate rather than hypothetical. Medical students in resource-constrained countries are using AI before many institutions have established formal guidance. In Uganda, 75.7% of 564 medical students reported having used AI tools, and 55.5% reported using them for assignments.5 In Pakistan, 61% of surveyed medical and dental students reported prior academic use of ChatGPT, commonly for case-based learning, problem-based learning, and clarification of topics.6 These findings support a shift from debating whether students should encounter AI to determining how institutions can make existing use safer, more educationally defensible, and more equitable. This commentary proposes a Somalia-centered framework for that task. AI is positioned as a supervised educational assistant that can extend access to explanation, deliberate practice, and feedback, but not as a substitute for teachers, clinical supervisors, patients, bedside learning, or professional judgment. The central contribution is an implementation logic that starts with local constraints and links each proposed AI use to a governance safeguard, a phased institutional action, and an outcome that can be evaluated rather than assumed. Commentary Approach and Framework Development This article is a commentary and does not report primary data, a systematic review, or an empirically validated intervention. The proposed framework was developed through a focused narrative synthesis of three evidence streams: Somalia-specific literature on health-professions education and workforce constraints; peer-reviewed literature on AI tutoring, virtual patients, feedback, assessment, learner use, and implementation risks; and international guidance on responsible AI in health and education.1–4,7–10 The synthesis was used to identify recurring implementation requirements and to translate them into a practical sequence of context, educational function, safeguard, implementation phase, and proposed outcome. Accordingly, the indicators presented in Table 1 and the pathway shown in Figure 1 are proposals for future evaluation rather than claims of demonstrated effectiveness in Somalia. This distinction is important because the educational impact, cost, equity effects, and safety of AI are likely to vary by institution, language, connectivity, faculty capacity, and clinical context. Table 1Proposed AI Applications, Safeguards, Supporting Evidence, and Outcomes for Medical Student Education in Somalia and Low-Resource SettingsEducational DomainSomalia/Low-Resource ConstraintProposed AI-Supported FunctionRequired Safeguard and Evidence BaseProposed Outcome for EvaluationFoundational sciencesVariable access to tutors and uneven conceptual support.Adaptive explanations, retrieval questions, bilingual clarification, and self-directed revision.Faculty-approved prompts and verification against trusted learning resources; AI tutoring should remain formative.2–4,11Knowledge gain and retention; self-directed learning; module performance.Clinical reasoningUnequal case exposure and limited standardized-patient or simulation capacity.Faculty-authored virtual patients for history taking, differential diagnosis, and repeated practice.Local case validation, human supervision, and structured feedback. Prospective and randomized studies support feasibility and selected learning effects.12,13History-taking quality; clinical reasoning scores; OSCE performance; diagnostic calibration.Assessment and feedbackLarge cohorts and limited faculty time for frequent formative feedback.Draft MCQs, short-answer prompts, OSCE stations, rubrics, and individualized feedback.Faculty review before release; no unvalidated AI scoring for high-stakes decisions. Model performance varies across professional examination tasks.14,15Feedback turnaround; item quality; learner performance; scoring agreement where evaluated.Multilingual communicationEnglish-based instruction with Somali patient communication and some Arabic use.Bilingual/trilingual explanations, role-play, and counseling practice.Faculty-verified terminology, back-checking of translations, and prohibition of unsupervised clinical translation for patient care.Terminology accuracy; semantic equivalence; communication performance; learner confidence.Research literacyLimited access to research mentorship and scientific writing support.Explanation of study design, question refinement, outlining, and language editing.Mandatory verification of claims and citations; prohibition of fabricated data, references, ethics approvals, or authorship.Proposal quality; methods knowledge; citation accuracy; responsible-use adherence.Faculty and curriculum supportHigh clinical workload and limited preparation time.First drafts of lesson plans, cases, checklists, and curriculum mapping.Faculty ownership and approval of final content; local guideline alignment.Preparation time; content accuracy; consistency of teaching materials.Academic integrityUnsupervised use for assignments and examinations may bypass authentic learning.Transparent AI-use declarations and supervised formative use.Approved-use matrix, oral defense, supervised assessment, OSCEs, and misconduct procedures.Disclosure compliance; academic-integrity events; authentic performance.Patient confidentialityRisk of entering identifiable clinical information into public systems.Use only anonymized educational cases or institutionally approved systems.No identifiable patient data in unapproved tools; data minimization and institutional governance.8,16Privacy incidents; student knowledge of confidentiality; compliance with policy.Equity and accessUnequal connectivity, devices, data costs, and access to premium tools.Mobile-friendly, low-bandwidth, institutionally supported access where feasible.Monitor digital exclusion and avoid making paid tools a prerequisite; broader LMIC implementation requires attention to infrastructure and trust.7,17Participation by subgroup; connectivity failures; cost per learner; access equity.Local health-system relevanceAI may recommend tests, medicines, or referrals that are unavailable or unaffordable.Faculty-adapted prompts and cases reflecting local disease patterns and service availability.Clinical validation against local protocols, medicines, diagnostics, and referral pathways.Appropriateness of recommendations; local feasibility; preparedness for practice.Abbreviations: AI, artificial intelligence; MCQ, multiple-choice question; OSCE, objective structured clinical examination. Figure 1Human-centered, Somalia-centered pathway for responsible AI integration in medical student education.The flowchart details AI integration in Somali education, starting with the ′Somalia-Specific Context′ of rapid education growth, urban focus, multilingual learning and mobile-first, low-bandwidth environment with 28% internet use. This context leads to three areas: ′Personalized Learning′ with adaptive explanations and bilingual support; ′Clinical Reasoning′ with virtual patients and case-based practice; and ′Faculty & Assessment′ focusing on teaching materials and research support. These areas connect to the ′Governance Gate′ involving faculty oversight, academic integrity and local clinical validation. The timeline has three phases: ′0-6 Months′ for policy assessment and AI literacy; ′6-18 Months′ for curriculum integration and equity monitoring; and ′18-36 Months′ for scaling validated uses and national coordination. Proposed outcomes include improved knowledge, clinical reasoning, safer AI use, equitable access, faculty efficiency and graduates responsive to local needs.A flowchart detailing AI integration in Somali education, covering context, learning, governance and outcomes. Table 1Proposed AI Applications, Safeguards, Supporting Evidence, and Outcomes for Medical Student Education in Somalia and Low-Resource SettingsEducational DomainSomalia/Low-Resource ConstraintProposed AI-Supported FunctionRequired Safeguard and Evidence BaseProposed Outcome for EvaluationFoundational sciencesVariable access to tutors and uneven conceptual support.Adaptive explanations, retrieval questions, bilingual clarification, and self-directed revision.Faculty-approved prompts and verification against trusted learning resources; AI tutoring should remain formative.2–4,11Knowledge gain and retention; self-directed learning; module performance.Clinical reasoningUnequal case exposure and limited standardized-patient or simulation capacity.Faculty-authored virtual patients for history taking, differential diagnosis, and repeated practice.Local case validation, human supervision, and structured feedback. Prospective and randomized studies support feasibility and selected learning effects.12,13History-taking quality; clinical reasoning scores; OSCE performance; diagnostic calibration.Assessment and feedbackLarge cohorts and limited faculty time for frequent formative feedback.Draft MCQs, short-answer prompts, OSCE stations, rubrics, and individualized feedback.Faculty review before release; no unvalidated AI scoring for high-stakes decisions. Model performance varies across professional examination tasks.14,15Feedback turnaround; item quality; learner performance; scoring agreement where evaluated.Multilingual communicationEnglish-based instruction with Somali patient communication and some Arabic use.Bilingual/trilingual explanations, role-play, and counseling practice.Faculty-verified terminology, back-checking of translations, and prohibition of unsupervised clinical translation for patient care.Terminology accuracy; semantic equivalence; communication performance; learner confidence.Research literacyLimited access to research mentorship and scientific writing support.Explanation of study design, question refinement, outlining, and language editing.Mandatory verification of claims and citations; prohibition of fabricated data, references, ethics approvals, or authorship.Proposal quality; methods knowledge; citation accuracy; responsible-use adherence.Faculty and curriculum supportHigh clinical workload and limited preparation time.First drafts of lesson plans, cases, checklists, and curriculum mapping.Faculty ownership and approval of final content; local guideline alignment.Preparation time; content accuracy; consistency of teaching materials.Academic integrityUnsupervised use for assignments and examinations may bypass authentic learning.Transparent AI-use declarations and supervised formative use.Approved-use matrix, oral defense, supervised assessment, OSCEs, and misconduct procedures.Disclosure compliance; academic-integrity events; authentic performance.Patient confidentialityRisk of entering identifiable clinical information into public systems.Use only anonymized educational cases or institutionally approved systems.No identifiable patient data in unapproved tools; data minimization and institutional governance.8,16Privacy incidents; student knowledge of confidentiality; compliance with policy.Equity and accessUnequal connectivity, devices, data costs, and access to premium tools.Mobile-friendly, low-bandwidth, institutionally supported access where feasible.Monitor digital exclusion and avoid making paid tools a prerequisite; broader LMIC implementation requires attention to infrastructure and trust.7,17Participation by subgroup; connectivity failures; cost per learner; access equity.Local health-system relevanceAI may recommend tests, medicines, or referrals that are unavailable or unaffordable.Faculty-adapted prompts and cases reflecting local disease patterns and service availability.Clinical validation against local protocols, medicines, diagnostics, and referral pathways.Appropriateness of recommendations; local feasibility; preparedness for practice.Abbreviations: AI, artificial intelligence; MCQ, multiple-choice question; OSCE, objective structured clinical examination. Proposed AI Applications, Safeguards, Supporting Evidence, and Outcomes for Medical Student Education in Somalia and Low-Resource Settings Abbreviations: AI, artificial intelligence; MCQ, multiple-choice question; OSCE, objective structured clinical examination. Human-centered, Somalia-centered pathway for responsible AI integration in medical student education. Context and Scientific Rationale Somalia’s educational context creates both a need for scalable learning support and a high threshold for responsible implementation. Rapid institutional expansion occurs alongside uneven access to specialist educators, simulation facilities, structured feedback, research mentorship, and standardized clinical assessment.1 AI cannot correct accreditation gaps, workforce shortages, or inadequate clinical placements. It can, however, provide additional supervised opportunities for explanation, practice, and feedback while those structural reforms continue. The scientific rationale is strongest where AI is used for educational augmentation rather than autonomous instruction. A prospective study of a GPT-4-powered simulated patient found medically plausible responses in more than 99% of 1,894 question-answer pairs and almost perfect agreement between AI and a human rater for overall history-taking feedback (Cohen kappa=0.832), although agreement was weaker in several feedback categories.12 In a randomized controlled trial, students who received LLM-generated feedback during simulated history taking improved clinical decision-making more than students who practiced with AI patients without feedback after four training sessions.13 Earlier randomized evidence in simulated surgical training also found higher performance with an AI tutoring system than with remote expert instruction, demonstrating that structured, metric-based AI feedback can support deliberate practice in defined tasks.11 These studies do not establish that AI is universally effective, but they provide a stronger basis than hypothetical benefit alone for carefully designed educational pilots. Assessment evidence likewise demonstrates both capability and limitation. In dental licensing examination questions, GPT-4.0 outperformed GPT-3.5 and passed the written thresholds used in the study, but performance varied substantially by model and examination.14 A later comparison of GPT-4.0, Claude-2, and Llama-2 on prosthodontic and restorative dentistry questions found marked differences in both answer accuracy and rationale quality.15 Such variability is directly relevant to medical education: fluent output should not be equated with stable expertise, and model performance in one domain should not be generalized to another without validation. Language adds a distinctive implementation challenge. Medical teaching in Somalia commonly uses English, while students and patients communicate primarily in Somali and may also use Arabic. Generative AI can assist with explanation and translation, but clinical terminology may be mistranslated, simplified incorrectly, or rendered in language that is culturally understandable but medically imprecise. For educational use, bilingual or trilingual AI outputs should therefore be verified against faculty-approved terminology and, where possible, back-checked between languages before they are incorporated into teaching materials, patient-counseling exercises, or assessment content. Evidence-Informed Educational Applications Personalized Learning and Tutoring AI can provide stepwise explanations, adaptive questioning, retrieval practice, and immediate clarification. The educational value is not simply that an LLM can generate more text; it is that the learner can request explanation at different levels, test understanding, and receive task-specific feedback. This function is most defensible when prompts are tied to explicit learning objectives and outputs are checked against trusted textbooks, national guidance, or faculty-approved material.2–4,11 Clinical Reasoning and Virtual Patients LLM-based virtual patients can give learners repeated opportunities to practice history taking, differential diagnosis, and information gathering when standardized-patient capacity is limited. Empirical studies show that simulated patient interactions can be plausible and that structured AI-generated feedback may improve selected clinical reasoning behaviors.12,13 In Somalia, such cases should be locally authored or faculty-adapted so that disease prevalence, available investigations, medication access, and referral pathways reflect the setting. AI-generated cases should supplement, not reduce, contact with real patients and clinical supervisors. Formative Assessment and Feedback AI can assist faculty in drafting multiple-choice questions, short-answer prompts, OSCE stations, rubrics, and individualized formative feedback. Licensing-examination studies demonstrate that advanced models can answer many professional assessment items correctly, but model-to-model variation and incorrect rationales remain important limitations.14,15 Faculty review is therefore required before AI-generated assessment material is released to learners, and high-stakes examinations should not depend on unvalidated AI scoring. Research Literacy and Academic Writing AI can explain study designs, help learners structure search questions, improve language clarity, and provide examples of reporting conventions. The educational objective should be research literacy, not automated authorship. Students must verify every factual claim and citation, disclose AI assistance where required, and never use AI to fabricate references, data, ethical approvals, patient histories, or authorship contributions. Faculty and Curriculum Support In settings where clinical teachers carry substantial service responsibilities, AI can reduce routine preparation time by generating first drafts of lesson plans, case discussions, checklists, and feedback templates. Such efficiencies are useful only when faculty retain ownership of the final content. As summarized in Table 1, each educational application should be paired with a specific safeguard and a proposed outcome that can be measured locally rather than presumed. Realistic Technology Options Near-term implementation does not require a bespoke national AI platform. Medical schools can begin with institutionally approved browser-based LLM interfaces, low-bandwidth learning-management-system assistants, or faculty-authored virtual-patient chat interfaces. Where technical capacity and governance permit, locally hosted open-source language models may later reduce dependence on commercial subscriptions and improve control over data, but they introduce additional requirements for hardware, maintenance, cybersecurity, and local validation. These categories are implementation options rather than endorsements of any specific product. Governance, Ethics, Equity, and Multilingual Safety The first governance priority is institutional accountability. Each medical school should place AI oversight within an existing academic structure, such as the curriculum or medical education committee, and include clinical faculty, educational leadership, information technology, ethics or research governance, and student representation. The group should maintain an approved-use matrix that distinguishes low-risk uses such as formative explanation from restricted uses such as assessment generation and prohibited uses such as entering identifiable patient data into unapproved public systems. Accuracy and hallucination remain central risks. Generative AI can provide incorrect drug doses, fabricated citations, plausible but false explanations, or recommendations that exceed local resources. Model performance also changes over time and differs across systems, as illustrated by licensing-examination studies.14,15 Institutions should therefore validate representative prompts and cases locally, require source checking for factual outputs, and create a simple mechanism for students and faculty to report unsafe or misleading outputs. Patient confidentiality requires a stricter boundary. Students should not upload identifiable histories, images, laboratory results, or other protected clinical information to public AI tools unless an institution has explicitly approved the system and data-processing arrangement. Ethical analyses of AI use among vulnerable populations emphasize consent, privacy, algorithmic bias, the digital divide, and the risk that technological convenience can widen existing inequalities.16 These concerns are highly transferable to fragile and low-resource health systems. Equity must also be designed into implementation. National internet use in Somalia was about 28% in 2024, illustrating the broader connectivity environment in which medical schools operate.17 Although medical students may have higher connectivity than the general population, reliable devices, electricity, mobile data, and paid subscriptions cannot be assumed. Institutions should therefore prioritize mobile-friendly, low-bandwidth access, shared institutional resources where feasible, and learning activities that do not penalize students who cannot use commercial premium tools. International guidance provides a useful ethical foundation but requires local operationalization. WHO emphasizes human oversight, transparency, safety, data protection, and bias mitigation for large multimodal models in health.8 UNESCO similarly recommends human-centered governance, privacy protection, institutional preparedness, and pedagogical oversight in education.9 The AAMC emphasizes responsible, equitable, transparent use, faculty development, privacy, interdisciplinary curriculum design, and continuous monitoring in medical education.10 For Somalia, these principles need to be translated into concrete institutional rules, faculty verification processes, multilingual quality checks, and equity safeguards rather than adopted only as high-level statements. Novelty and Contribution of the Somalia-Centered Framework The originality of the proposed framework is not a claim that its ethical principles are new. Human oversight, privacy, transparency, and equity are already established in international guidance.8–10 The additional contribution is to operationalize those principles for a setting where four constraints occur together: rapid expansion and variable standardization of health-professions education; limited faculty and simulation capacity; a multilingual Somali-English-Arabic learning environment; and mobile-first, low-bandwidth access with geographic inequity.1,17 The framework therefore differs from a generic AI ethics checklist in three ways. First, it begins with local educational constraints rather than with technology capabilities. Second, it requires every AI-supported function to pass through a governance gate that addresses faculty verification, academic integrity, confidentiality, translation safety, and local clinical validity. Third, it connects adoption to a phased implementation roadmap and explicit educational outcomes. Figure 1 illustrates this systems-oriented sequence from Somalia-specific context to educational function, safeguards, implementation phase, and proposed outcomes. The framework remains conceptual; its added value will ultimately depend on whether local institutions can test, refine, and validate it. Phased Implementation Roadmap for Somali Medical Schools Short Term (0–6 Months): eStablish Governance and Low-Risk Pilots Institutions should conduct a baseline assessment of student and faculty AI use, designate an accountable AI education group, publish an approved-use matrix, and provide basic AI-literacy training on hallucination, verification, privacy, academic integrity, and prompt design. Initial pilots should be limited to low-risk functions such as formative concept explanation, faculty-reviewed question generation, or supervised virtual history-taking cases. Baseline measures should include access, current use, perceived confidence, faculty workload, and documented errors. Medium Term (6-18 Months): Integrate Validated Uses Into Curriculum and Assessment Uses that perform acceptably in pilot evaluation can be mapped to learning outcomes in medical informatics, research methods, evidence-based medicine, clinical skills, and ethics. Somali-English-Arabic terminology sets should be faculty-validated, and AI-supported cases should be checked against local disease patterns, available investigations, medicines, and referral pathways. Assessment reform should emphasize supervised writing, oral defense, OSCEs, workplace-based assessment, direct observation, and transparent AI-use declarations so that competence remains demonstrable even when AI is widely available. Long Term (18-36 Months): Scale, Evaluate, and Coordinate Institutions should scale only those applications that show educational benefit without unacceptable safety, equity, or integrity trade-offs. Multi-institutional evaluation can compare implementation across urban and less-resourced settings, examine cost and data requirements, and support shared standards. Where technical capacity permits, locally hosted or institutionally controlled systems may improve governance and reduce exposure of sensitive information. National regulators and medical schools can then use locally generated evidence to inform accreditation expectations and shared guidance rather than importing policies developed for different resource environments. Future Research Priorities Research should begin with descriptive and implementation questions before assuming effectiveness. Multi-institutional cross-sectional studies can establish how Somali medical students and faculty currently use AI, which tools are accessible, what training they have received, and where risks such as unverified clinical advice or patient-data entry occur. Qualitative interviews and focus groups can examine language, trust, cultural fit, faculty concerns, and the reasons students choose AI over conventional learning resources. Educational effectiveness should then be tested prospectively. Randomized or quasi-experimental studies could compare faculty-verified AI tutoring with usual self-study, or AI-supported virtual patient practice with standard case discussion. Appropriate outcomes include knowledge acquisition and retention, clinical reasoning scores, OSCE performance, accuracy of differential diagnosis, quality of history taking, learner satisfaction, self-directed learning, and the ability to identify AI errors. Multilingual studies should measure terminology accuracy and semantic equivalence across Somali, English, and Arabic rather than relying on subjective fluency alone. Implementation research should measure adoption, fidelity to institutional policy, frequency and type of AI errors, faculty verification time, connectivity failures, cost per learner, differential access by socioeconomic or geographic background, privacy incidents, and academic-integrity events. Stepped-wedge or cluster designs may be appropriate where schools introduce policy or curriculum changes at different times. These measures would allow future versions of the framework to be based on demonstrated educational value, feasibility, and equity rather than on technological enthusiasm. Conclusion Artificial intelligence is already entering medical students’ learning environments, including in resource-constrained countries, while institutional policy and faculty preparation often lag behind student adoption.5,6 For Somalia, the relevant question is therefore how to govern and evaluate AI use in a health-professions education system that has expanded rapidly while continuing to face uneven faculty, simulation, assessment, and digital resources.1 This commentary proposes a context-sensitive implementation framework that advances beyond a general statement of AI opportunities by linking Somalia-specific constraints to defined educational functions, mandatory safeguards, phased institutional actions, and measurable outcomes. Successful integration will require simultaneous investment in faculty development, multilingual verification, assessment reform, privacy protection, local clinical validation, and equitable access. AI should remain an adjunct to human teaching and clinical apprenticeship. If the proposed framework is prospectively tested and refined through local evidence, it may provide a practical route for using AI to expand supervised learning opportunities while preserving professional judgment, patient safety, and educational integrity in Somalia and comparable low-resource settings.