How Do AI Safety Events Get Reported?

In this week’s episode of the Health Data Ethics Podcast, I take on the two AI questions I hear often: what about the environmental costs, and what about job loss? Pope Leo XIV’s Magnifica Humanitas helped me shape answers to both, starting with his point that AI is never morally neutral and that all of us share responsibility for what it costs and what it gives.I also get into why opting out of AI doesn’t shrink its footprint, and what happened to radiology after Geoffrey Hinton told everyone to stop training radiologists.

Health Data Ethics Podcast is back with a new series on the moral and economic foundations of AI governance. In this episode, we're looking at Pope Leo XIV’s encyclical Magnifica Humanitas and the question at its center: what are we building?Most of the cultural conversation about AI treats its direction as inevitable, whether you’re hearing hype or doom. This letter insists we each have a responsibility to shape both the ends and the means, which matches the work I see every day.

This week's episode of the Health Data Ethics podcast tackles Education and Training, Element 7 on the Joint Commission / CHAI responsible AI framework. None of the first six elements do much if the people using the tool every day don't understand what it does, what it can't do, and what to do when something seems off, now and in the future.A few other items from this week's episode:End users need to know the difference between an AI that sounds confident and one whose reasoning can actually be trusted. How do your users know your AI system's limits well enough to know when to push back, just like they would on a colleague? How we can use our ongoing CME/CNE framework to do continued AI training, but only where it's most necessary.

In this episode, we tackle Joint Commission and CHAI framework element six: AI bias in healthcare. We talk about how bias enters through biased data, biased algorithms, and biased human actions on outputs. We look at the Obermeyer 2019 Science paper, the Joint Commission/CHAI RUAIH framework, and a striking convergence: Pope Leo XIV and economist Daron Acemoglu, from completely different traditions, reach the same conclusion. Ethics can't be a compliance checkbox at the end of AI development. It has to be built in from the start.

In healthcare, we have decades of safety event infrastructure built to help us collectively learn from failure. But when an AI tool makes an unsafe recommendation, degrades after a model update, or produces biased outputs, there's no systematic place for information to go, so that we, and others, can learn.That's the gap Joint Commission and Coalition for Health AI are starting to close with their Responsible Use of AI in Healthcare guidance.In this Health Data Ethics episode, I break down Element 5: voluntary, blinded reporting of AI safety events. What it means, how to use the structures you already have, and why the voluntary model supports more effective AI adoption in healthcare.

Your EHR vendor can push an update to the AI model inside your clinical decision support tool, and you might not even know it happened. What you validated last quarter may not be what you're running today, but you're still the one holding the liability.That's what we're covering in this Health Data Ethics Podcast episode, back after a summer break. We're talking about element four of the Joint Commission and CHAI guidance on AI in healthcare: ongoing quality monitoring. Vendors decide when to retrain a model, so the deploying health system has to be the one watching what happens to its own patients in its own environment.

I'm taking a brief break as work and life have both ramped up, leaving me in need of a few weeks to prep some great content for you. I'll be back later in the summer!

This week's Health Data Ethics podcast continues our series on the Joint Commission and CHAI guidance on the responsible use of health AI. In this episode we're digging into privacy and transparency.The guidance itself is reasonable. What I spent most of the episode on is how you actually implement it, because that's where things get interesting.Adding AI language to the Notice of Privacy Practices is a good first step, and a lot of health systems are doing it. But I think the most-told lie in modern life is still "I have read and agreed to the terms and conditions." Broad disclosure is honest, and it matters, and it's also not going to carry the whole weight of a transparent relationship with your patients.The piece I really wanted to dig into is opt-outs. If you offer patients the ability to opt out of something you can't actually turn off, you've built opt-out theater, and that erodes trust faster than just being honest about the limitation would. Ambulatory scribe is a real opt-out. Inpatient sepsis prediction is not technically feasible to opt out of, and we probably shouldn't pretend it is.I also spend some time on the clinician side, which I think gets short shrift in a lot of these conversations. Operational training on a tool is not the same thing as understanding how the model behaves, where it fails, and which patients it might be wrong for. Clinicians are the ones carrying accountability for human-in-the-loop judgment, and they need real explainability to do that well.

Getting an AI policy approved in a large health system is a different skill than writing one.In part two of my AI policy series on the Health Data Ethics Podcast, I share what months of drafting, socializing, and navigating formal approval at Cleveland Clinic actually looked like: the champions you need, the scope battles you'll face, and why the approval process is won or lost long before the policy enters formal review.The biggest takeaway: identify domains where your scope overlaps with someone else's, and get those leader in the room early before formal review even starts.

Writing an AI policy sounds straightforward — until it becomes the place where everyone in your organization hangs all their hopes and dreams for AI governance.In this episode of the Health Data Ethics Podcast, I walk through the first item on the Joint Commission and Coalition for Health AI's responsible use guidance: establishing an AI policy as your governance foundation. I share what we learned working on Cleveland Clinic's AI policy in late 2024 — before the JC/CHAI guidance even existed — including the structural traps that slow policies down, and why pre-approval stakeholder alignment is so important. If you're starting from zero or trying to get a stalled draft across the finish line, this one's for you.