University of Oxford

As a scientist and public speaker at the forefront of AI for health across research, practice, and policy, I firstly welcome the Global Dialogue on IA Governance and congratulate the Chairs on this endeavour for the UN to move abreast with the digital age. I am a Professor of Health Informatics and Biomedical Data Sciences at the University of Oxford as well as a Wellcome Trust AI Accelerator Fellow, and Principal Investigator of the Oxford PHI Lab. Our work focuses on leveraging fair and safe AI for equitable for health. To ensure the Dialog is successful, we must establish as an outcome a technical and regulatory baseline that ensures AI remains a "force for good" for both people and the planet. However there are some fundamental barriers, which may broadly be thought of in terms of a) "fit-for-use" technology (safety, effectiveness, trust), b) capacity /infrastructure, and c) governance/regulation. AI is not a monolith. It is imperative that the Dialog by design caters to the distinctive types of AI (e.g. classical v genAI) and their differing safety, ethical, regulatory, commercial, and environmental ramifications, in order for it to be meaningful. To be meaningful, the Dialogue must address AI's distinct sub-strata (e.g., Classical vs. GenAI) and their unique safety, ethical, and environmental ramifications. I propose three pillars for success: 1. Clinical Rigor & Environmental Accountability In healthcare, safety and effectiveness are mandates. Since regulators like the FDA and MHRA already classify AI as a medical device, the sector offers a blueprint for rigorous, outcome-based governance. Furthermore, we must address the carbon footprint of High-Performance Computing (HPC). Success requires balancing the energy demands of advanced analytics with sustainable infrastructure to protect the planet we aim to heal. 2. From Access to Agency for the Majority World Bridging the "AI divide" requires local ownership, agency, and digital upskilling. LMICs must not be mere consumers of Global North data; they need the agency to fine-tune models to regional health needs and linguistic nuances. This prevents "digital colonialism" and ensures AI is "fit-for-purpose" across diverse ecosystems. Secondly, levelling up of the technology industry is required both economically and accountability. This is where effective regulation is imperative. 3. Red Lines: From Bias to Super-intelligence We must transition from voluntary guidelines to auditable standards. This includes mandatory bias audits to protect marginalized groups and proactive frameworks to manage the trajectory toward super-intelligence. Establishing safeguards today ensures that as AI capabilities scale, they remain aligned with human agency and equity. Ultimately, success requires a multilateral, privacy-preserving equitable ecosystem that ensures AI to be leveraged for people and planet with safety, fairness, transparency, trust and accountability—ensuring technology augments human wisdom rather than replacing it. I would be happy to contribute further and participate in the in-person Dialog in Geneva as a speaker. Thank you Associate Professor Sara Khalid B.E., MSc. (Oxon.), D.Phil.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Open-source software, open data and open AI models

Based on the intersection of healthcare, climate action, and informatics, the following four thematic areas reflect the highest priorities for urgent action: 1. AI capacity-building 2. Social, economic, ethical, cultural, linguistic and technical implications of AI 3. Protection and promotion of human rights 4. Open-source software, open data and open AI models Explanation of Selection • AI capacity-building: Urgency lies in shifting from "digital colonialism" toward local agency. For LMICs, capacity-building must go beyond infrastructure to include the technical expertise required to link complex climate and health datasets for real-world adaptation, such as early warning systems. • Social, economic, ethical, cultural, linguistic and technical implications of AI: AI is a dual threat multiplier. We must urgently address the "market failure" of AI's environmental cost-such as the high carbon footprint of model training-and ensure ethical frameworks prioritize planetary health alongside economic gains. • Protection and promotion of human rights: All human rights must be protected throughout the AI lifecycle. In health, this means ensuring that shifting demographics-specifically the youngest and oldest generations-are not excluded or exploited by algorithmic biases that "bake in" health data poverty. • Open-source software, open data and open AI models: Data democratization is the "Achilles' heel" of equitable AI. Promoting open-source and open-data frameworks is essential for transparency and allows regional stakeholders to perform "fit-for-use" local validation, ensuring models understand regional medical nuances and specific health needs. By focusing on these areas, the Global Dialogue can ensure that AI governance transitions from high-level principles to actionable, rights-based frameworks that protect both people and the planet.

One critical cross-cutting issue currently missing is the systemic integration of Planetary boundaries into AI Governance. While the current themes address societal and ethical implications, they overlook the "silent" environmental cost of AI and its cascading impact on global health. My work at the Oxford Planetary Health Informatics (PHI) Lab-which uniquely bridges the gap between biomedical data science, engineering, and environmental epidemiology-demonstrates that AI governance cannot be siloed from planetary boundaries. 1. The AI-Climate-Health Nexus AI is a dual threat multiplier. The carbon footprint of training large-scale models currently rivals major industrial sectors, yet this cost is often borne by the very Global South communities that are most vulnerable to climate-driven health crises. Governance must include mandatory Environmental Sustainability Disclosures for AI systems to prevent "digital carbon debt" from undermining planetary health. 2. Data Sovereignty, Digital Monopoly, Local Validation The "AI Divide" is not merely about access to hardware; it is about the sovereignty of local expertise. Our recent research on fine-tuning Large Language Models for South Asian clinical notes shows that "off-the-shelf" models from the Global North often fail to capture regional medical nuances. Success requires a shift toward Local Fine-Tuning and Federated Analytics, ensuring that LMICs have the agency to govern their own data through frameworks like the OMOP Common Data Model, which we utilize to maintain privacy while scaling global insights. 3. Intergenerational Equity in a Shifting World Current rights-based themes lack a focus on shifting demographics. As we address AI's role in healthcare, we must consider the "digital detriment" youngest populations and the rapidly aging global population. Governance must ensure that AI serves as a tool for inclusive longevity, rather than a driver of diagnostic exclusion. By integrating these planetary and contextual dimensions, the Dialogue can move from abstract ethics to a practical, sustainable framework that protects both people and the planet.

In the healthcare and research sectors, the governance gap between rapid AI advancement and regulatory oversight creates a "safety-innovation paradox." While AI offers a solution to under-resourced fragile public health, the lack of standardized frameworks for local contextual validation remains a primary challenge. Significant Challenges • The "Black Box" Liability: In regions like the Global South, "off-the-shelf" models trained on Global North datasets often exhibit significant algorithmic bias. Without mandatory requirements for fine-tuning on local epidemiological data, we risk scaling "digital colonialism" and diagnostic inaccuracies that disproportionately affect marginalized groups. • Infrastructure & Literacy Gaps: A critical challenge is the disparity in digital literacy and High-Performance Computing (HPC) access. This creates a fertile ground for technology monopolies, where local healthcare systems become dependent on proprietary, non-transparent tools that may not align with regional health priorities and contexts. • Intergenerational Risks: Challenges posed by a global demographic shift require a delicate balance between youth bulges and digitally damaged childhoods and related physical and mental health challenges on one hand, and ageing populations and the opportunity for digitally aided self-care. Significant Opportunities • Regulatory Leadership: Healthcare regulatory bodies already views AI as a medical device (FDA/MHRA). This provides a unique opportunity to export clinical rigor—including mandatory bias audits and peer-reviewed safety standards—to other sectors.