7th Global Gastroenterology and Artificial Intelligence Summit (Live)

2- Shung

Pdf Summary

The presentation “From Bedside to Baseline: Shaping Foundation Models with Clinical Insight” by Dr. Dennis L. Shung emphasizes the critical role of robust evaluation, trustworthy value alignment, and human-AI co-evolution in deploying foundation AI models safely and effectively in clinical gastroenterology and hepatology.<br /><br />Key points include the variability in foundational model performance, with surveys indicating up to 60% inaccuracy of GPT-4 responses in GI/hepatology topics and widespread, though cautious, AI adoption by physicians. Rigorous evaluations—including agentic workflows evaluation and automated human-in-the-loop support—are essential to detect performance degradation and ensure patient safety. Trustworthy AI adoption depends on value alignment with clinician mental models so AI acts consistently with medical ethics and clinician judgment.<br /><br />Emergent challenges such as sycophancy (AI over-agreeing with users at the cost of accuracy) and misalignment with human preferences, especially concerning mental health, require careful attention. Government intervention and research into access, control, and risk management are necessary to mitigate catastrophic AI risks.<br /><br />Generative AI is being framed as a digital workforce revolutionizing healthcare workflows. Integrating AI into clinical care demands embedding values like helpfulness, honesty, and harmlessness, guided by medical ethical principles akin to the Hippocratic Oath. Practical strategies include prompt engineering to explicitly encode constraints, use of controlled vocabularies, and continuous contextual updates.<br /><br />Human-AI teaming aims to enhance clinical decision support with improved ease of use compared to prior models, driving efficiency gains in guideline development and care delivery. The co-evolution of clinicians and AI technology is portrayed as critical: clinicians must lead the design of meaningful evaluations, codify value alignment in AI behavior, and direct development to focus on patient-centric outcomes. This approach promises to elevate patient care by blending human insight with AI capabilities while maintaining safety and trust.

Keywords

Foundation Models

Clinical Insight

Gastroenterology

Hepatology

AI Evaluation

Value Alignment

Human-AI Co-evolution

Generative AI

Medical Ethics

Patient Safety