Key Details
The ILO argues that access to AI does not automatically improve care. The gains depend on how AI is integrated into clinical workflows.
Aspect | Finding | Why It Matters |
|---|---|---|
Randomized trial | 249 physicians in Indonesia, Kenya and the Netherlands completed clinical cases with or without GPT-4o assistance. | Provides robust causal evidence rather than observational findings. |
Clinical performance | AI-assisted physicians achieved higher scores in all three countries, with the largest gains in Kenya (+18%). | Suggests AI may help reduce capability gaps in resource-constrained settings. |
Human–AI interaction | The best-performing physicians using AI outperformed both the AI alone and physicians without AI. | Indicates AI complements clinical expertise rather than replacing it. |
Main risks | Automation bias, hallucinations, inappropriate recommendations and poor alignment with local practice. | Demonstrates that unsafe integration can reduce decision quality. |
Governance priorities | Local validation, AI literacy, workflow redesign, liability frameworks, monitoring and social dialogue. | Shows that institutional design determines whether AI improves healthcare. |
Summary
The Experiment Shows That AI Alone Is Not the Answer
Unlike many discussions based on simulations or benchmark tests, this paper evaluates AI through a randomized controlled trial involving 249 physicians across Indonesia, Kenya and the Netherlands. Doctors were randomly assigned to complete standardised clinical cases either with or without access to GPT-4o.
Physicians using AI consistently performed better, improving clinical reasoning scores by 18 percentage points in Kenya, 10.7 points in Indonesia, and 7.2 points in the Netherlands. The largest improvements occurred in resource-constrained settings and among non-specialist doctors, suggesting that LLMs can help narrow differences in clinical capability where specialist expertise is limited.
Yet the experiment also delivers a more important message: providing access to AI does not automatically improve clinical decisions. Some physicians performed worse despite AI assistance, indicating that technology alone cannot guarantee better care.
The Greatest Gains Come from Combining Clinical Judgement with AI
One of the study’s most significant findings is that AI performs best alongside experienced clinicians rather than instead of them.
The LLM on its own outperformed many physicians who worked without AI support. However, the highest-scoring physicians who used AI exceeded the performance of both the standalone AI model and their peers working without AI. This suggests that experienced clinicians are better able to interrogate AI outputs, identify weak recommendations and integrate machine-generated insights with clinical judgement.
The findings therefore support an augmentation model, where AI strengthens diagnostic reasoning, differential diagnosis and case analysis, rather than a substitution model in which AI replaces physicians.
Poor Integration Can Increase Clinical Risks
The study finds that the effectiveness of AI depends less on the quality of the model than on the way clinicians use it.
Several physicians accepted AI recommendations too readily, illustrating automation bias—the tendency to over-trust machine-generated advice even when it is incorrect. Participants also identified hallucinated responses, inaccurate recommendations and suggestions that failed to match local treatment protocols or resource availability.
These findings indicate that safe deployment requires structured workflows that preserve independent clinical reasoning, require verification of AI outputs and prevent clinicians from becoming passive reviewers of machine-generated advice.
The report also emphasises that AI literacy extends beyond learning to operate AI tools. Physicians need skills to critically evaluate outputs, recognise hallucinations, frame effective prompts and understand the limitations of AI-generated recommendations.
Governance, Not Technology, Will Determine Long-Term Success
The paper argues that AI adoption in healthcare is fundamentally a governance challenge.
It recommends validating AI models against local clinical guidelines, disease patterns and health-system realities before deployment, particularly because models trained predominantly on high-income-country data may recommend investigations or treatments that are unavailable or inappropriate elsewhere.
The report also calls for continuous monitoring of AI-assisted care, clear legal frameworks defining liability for AI-supported decisions, and institutional mechanisms that involve healthcare professionals in designing and governing AI deployment. It further raises concerns about digital sovereignty, arguing that excessive dependence on foreign commercial AI models could limit national control over healthcare systems and future innovation.
Rather than asking whether AI is sufficiently capable, the paper concludes that policymakers should focus on designing governance systems that ensure AI consistently supports, rather than compromises, clinical decision-making.
What is Clinical Reasoning?
Clinical reasoning is the cognitive process through which healthcare professionals gather patient information, interpret symptoms and test results, evaluate possible diagnoses and determine appropriate treatment. It combines medical knowledge, experience and professional judgement. AI can assist this process by analysing information and suggesting possible diagnoses or management options, but responsibility for clinical decisions remains with the physician.
Policy Relevance
As India expands the Ayushman Bharat Digital Mission, AI should be deployed as a clinical decision-support capability integrated into existing healthcare workflows rather than as an autonomous diagnostic tool.
Institutions such as the National Medical Commission (NMC) and Indian Council of Medical Research (ICMR) can embed AI literacy into medical education, including prompt design, critical appraisal of AI outputs, recognition of hallucinations and safeguards against automation bias.
AI systems should be validated against Indian clinical guidelines, disease patterns, languages and resource constraints before deployment to ensure recommendations remain clinically relevant across diverse healthcare settings.
India should establish clear accountability and liability frameworks governing AI-assisted medical decisions, alongside continuous post-deployment monitoring and reporting of safety outcomes.
The report also highlights the strategic importance of indigenous health AI. Strengthening locally developed and locally governed AI models under initiatives such as the IndiaAI Mission can reduce dependence on external platforms while improving alignment with India’s healthcare priorities.
Given India’s shortage of specialist healthcare professionals, AI has the potential to improve clinical decision-making in underserved regions, provided deployment is accompanied by robust governance, workforce preparedness and digital infrastructure.
Follow the Full Paper Here: Does a General-Purpose Large Language Model Improve Physicians’ Clinical Reasoning?