AI Rounds by the Cumming School of Medicine

Delegating to an AI agent is not the same as prompting a chatbot. In this episode, I walk through what good delegation actually looks like — scope, constraints, and checkpoints — and apply it to three concrete tasks: updating your CV, building a simple course assistant, and screening abstracts for a scoping review. The framework is practical, the failure modes are real, and the limits matter.

Agents went from a concept most faculty hadn't encountered to embedded infrastructure in the tools you already use — in roughly 18 months. In this episode, I trace what changed, why it changed so fast, and what a three-tier framework for thinking about AI autonomy can tell you about what you're already working with.

Microsoft & Health Management Academy: Agentic AI readiness in healthcare — 43% piloting, 3% deployed in live workflows. NEJM, January 2026. Cao W, Zhang Q, Liu J, Liu S. From Agents to Governance: Essential AI Skills for Clinicians in the Large Language Model Era. J Med Internet Res. 2026;28:e86550.

Most conversations about AI in medical education focus on efficiency — faster feedback, streamlined assessment, reduced administrative burden. Dr. Nia Abdullayeva asks a harder question: what happens to the human dimensions of training in the process? In this episode, Nia joins me to explore how AI, used intentionally, can protect rather than displace the relational time that makes good teaching and good medicine possible. We get into the cognitive overload driving compassion fatigue in learners, how AI can support sustainable feedback practices, and why the hidden curriculum in medical training might be one of the places AI has the most to offer. Whether you work in the clinic, the classroom, or the research lab, Nia's perspective on AI and professional identity is one I think will stay with you.

Not all AI tools work the same way — and NotebookLM is a clear example of why that distinction matters. In this episode, I walk through how NotebookLM works, why its underlying architecture makes it meaningfully different from general-purpose AI assistants, and what that means for how you should use it. We cover the full range of what the tool can produce — from cited Q&A and study guides to AI-generated audio overviews, infographics, and slide decks — and I spend real time on use cases for educators, a group I think has been underserved in some of our earlier conversations. We also get into the privacy and data governance considerations that anyone in academic medicine needs to understand before uploading anything to a cloud-based AI tool. If you've been curious about NotebookLM but weren't sure where it fits in your work, or whether it's appropriate for your context, this episode is your starting point. This episode also serves as an early introduction to Retrieval-Augmented Generation — a concept we'll return to in depth later this season.

https://notebooklm.google

AI agents are everywhere now — tools that can manage your email, execute research tasks, book appointments. But before we explore what AI agents can do for you, we need to understand what you're actually trading every time you use AI.

In this Season 2 opener, I walk through the major trade-offs faculty face when using AI in research, education, and clinical work: efficiency versus depth, breadth versus expertise, automation versus agency, and convenience versus privacy. Using the same risk/benefit framework you apply to clinical decisions, I'll help you evaluate when AI use makes sense and when the trade-off isn't worth it.

Understanding these fundamentals — especially around agency and privacy — becomes critical as AI takes on more autonomous roles. You're already skilled at risk/benefit analysis. You just need to apply it to a new domain.

https://www.media.mit.edu/publications/your-brain-on-chatgpt/

This brief episode wraps up Season 1 of AI Rounds. Thank you, listeners, for tuning in, we've covered a lot of ground together — from how AI models learn to practical applications in teaching, research, and clinical work. Season 2 will launch in Spring 2026, and I'd love your help shaping upcoming content. If you have topic suggestions, please send them to [email protected].

In an era where AI tools promise to accelerate every aspect of academic work, graduate students face a paradox: having access to powerful technology while needing to develop fundamental research skills.

In this episode, Inara Lalani, a current graduate student shares insights about the critical importance of discernment in AI use. The conversation explores prioritizing process over product in the learning environment, developing frameworks for deciding when AI helps versus hinders their learning, cultivating critical thinking skills that will serve them throughout their research careers, and unexpected ways GenAI is impacting graduate research.

Join us for a nuanced and thought provoking conversation that touches on several themes we've seen throughout this season.

What if your AI could work independently toward your goals instead of just answering individual questions? In this episode, we explore AI agents—autonomous systems that can monitor information, make decisions, and take actions without constant oversight.

Unlike traditional AI tools that respond to single requests, agents operate continuously to achieve specific objectives. For medical faculty, this means AI that can monitor research literature, track administrative deadlines, support educational workflows, and enhance clinical decision-making.

In this episode we break down what AI agents actually are, explores types most relevant to clinical practice and medical education, and provides a practical framework for building your first agent using no-code platforms. The episode covers essential considerations for medical environments, including privacy, security, and integration with existing systems.

Links from this episode:

zapier.com

ifttt.com

microsoft.com/en-us/power-platform/products/power-automate

Feeling overwhelmed by AI in education? You've been here before.

In this conversation, Dr. D'Arcy Norman draws on three decades of educational technology experience to reveal a striking pattern: roughly every ten years, a "revolutionary" technology emerges that promises to transform education forever. Computers. The Internet. MOOCs. And now, AI.

Each time, the same fears surface. Each time, vendors promise disruption and personalization. And each time, education evolves—not by replacing human connection, but by thoughtfully integrating new tools into teaching practice.

D'Arcy shares insights from his work leading learning technology initiatives at the University of Calgary, offering medical educators a practical perspective for evaluating AI tools while preserving what matters most: the relationships, mentorship, and clinical judgment that form the core of medical training.

If you're a medical faculty member navigating AI fears, wondering how to maintain academic integrity, or simply trying to understand where AI fits in your teaching, this episode provides both historical perspective and actionable guidance. The message is clear: teachers won't be replaced by AI, and student learning won't be diminished—provided we embrace intentional course design and authentic assessment.

The wave will pass. The question is how we ride it.

Why do GenAI systems confidently state incorrect medical facts instead of saying "I don't know?" Groundbreaking research from OpenAI and Georgia Tech reveals that AI hallucinations aren't bugs to be fixed—they're inevitable consequences of how these systems are trained. This episode explores the "singleton problem" that makes AI systematically unreliable on rare facts, connects to our previous discussion of AI benchmark saturation (Episode 9), and explains why the same evaluation methods that create impressive test scores actually reward confident guessing over appropriate uncertainty. For medical faculty evaluating AI tools, understanding these statistical realities is crucial for teaching students, conducting research, and developing institutional policies that account for AI's fundamental limitations.

Links from this episode:

https://openai.com/index/why-language-models-hallucinate