Discuss the ethical challenges of large multi-modal AI models in healthcare. How far do the ICMR (2023) and WHO (2024) frameworks address them?

Large multi-modal models (LMMs) are generative AI systems. They take in text, images and other data and produce human-like outputs for diagnosis, clinical care and patient queries. WHO's January 2024 guidance names their risks [1]. The ICMR and WHO frameworks identify these risks well. They do less to make them enforceable.

Ethical challenges of LMMs

• Harmful outputs: they can give false, incomplete or biased statements that mislead health decisions [1].
• Bias and equity: they may be trained on poor-quality or biased data, and the best models may be unaffordable, which can widen the digital health divide [1].
• Automation bias: clinicians may miss errors or hand difficult decisions over to the machine [1].
• Privacy and cybersecurity: patient data can be exposed [1], and consent is hard to make meaningful for a patient who cannot read the form [4].
• Accountability: it is unclear whether the developer, the hospital or the doctor answers for an AI error.

ICMR Guidelines (2023): how far?

• (+) They set out ten patient-centric principles, including accountability and liability, autonomy, data privacy, non-discrimination and validity [2].
• (+) Under autonomy, they require human oversight and let patients refuse AI-based care [2].
• (–) They were written for AI in general, before LMMs spread widely. ICMR calls them a "living document" that still has to be updated [2].
• (–) They are advisory: no licence, penalty or recall power. Only data processing is legally binding, through the DPDP Act, 2023 [3].

WHO Guidance (2024): how far?

• (+) It gives 40+ recommendations for governments, developers and providers [1].
• (+) It calls for mandatory independent post-release audits, with results broken down by age, race and disability, and for involving patients from the design stage [1].
• (+) It builds on WHO's 2023 call for representative datasets and external validation [4].
• (–) It is non-binding. Governments are asked to name a regulator only "as resources allow" [1], and liability is left unsettled.

Assessment: both frameworks cover the ethical risks broadly, but they enforce very little.

Together, the two frameworks map the ethical risks of LMMs fairly fully but turn few of them into legal duties. India should update the ICMR guidelines for generative AI and require audits before and after deployment. It should also set out liability clearly while keeping clinicians in the loop. That would protect the rights to health and privacy under Article 21 and advance SDG 3 without giving up AI's diagnostic promise.