Five TEDx Lessons on AI, Medicine, and Human Judgment
What five TEDx stages taught an emergency physician about explaining AI to non-technical audiences, and what those lessons reveal about clinical judgment itself.
I have given five TEDx talks. Each one was harder than the last, which is not how I expected that to go.
The difficulty was never the technology. It was the compression. A TEDx stage gives you a short window, no jargon protection, and an audience that will politely disengage the moment you start sounding like a conference panel. Preparing for that repeatedly taught me things about AI in medicine that I did not learn from writing books about it.
My talks have covered AI and healthcare, safe AI, edge technology, satellites, emergency medicine, preventive care, and the future of clinical judgment. The topics sound different from the outside. Underneath, they keep returning to the same question that sits at the center of DR GPT™: how do we use powerful tools without losing the human responsibility that makes medicine medicine?
Here are the five lessons that stuck.
One: if you cannot explain it without the vocabulary, you do not understand it well enough to deploy it
Healthcare AI has an enormous amount of language designed to make ordinary ideas sound proprietary.
On a TEDx stage that language dies immediately. You cannot say "multimodal inference pipeline" to a general audience and survive. You have to say what the thing does, to whom, and what happens when it is wrong.
That constraint turned out to be a governance tool. Now, when a vendor presents to a hospital committee, I ask them to explain the product the way they would explain it to a room of people who do not work in healthcare. The ones who can do it usually have a real product. The ones who cannot are often hiding a thin one behind vocabulary.
Try it at your next demo. It works better than any technical question you could ask.
Two: the story has to be about a person, or nobody remembers the number
I can tell an audience that a widely deployed sepsis model showed an area under the curve of 0.63 and missed 67 percent of cases in external validation, which is Wong A, Otles E, Donnelly JP, et al., JAMA Internal Medicine, 2021, per PubMed (DOI).
That statistic is important. It is also forgettable on its own.
What people remember is the nurse. The one who has silenced that alert nine hundred times, who now moves past it without reading it, and who is doing something completely rational that the system has trained her to do.
Same fact. One version changes how a board votes.
That is why I keep bringing healthcare AI back to the bedside. A patient does not experience an algorithm as a model score. A patient experiences a delay, a missed signal, a rushed explanation, a better handoff, or a clinician who finally has time to listen. If the story does not reach that level, it is not ready for healthcare leadership.
Three: audiences are not afraid of AI, they are afraid of being handled
Every audience I have faced arrives with a version of the same question underneath whatever they actually ask. Will someone be accountable to me, or will I be handed a decision made by a system nobody will explain?
That fear is not irrational and it is not technophobia. It is a governance concern in emotional clothing.
Which is why the answer that lands is never reassurance about how accurate the model is. It is a description of who is accountable, what happens when the system is wrong, and how a person can push back. Audiences relax when they hear a name and a process. They tense up when they hear a percentage.
Legislators have been converging on the same instinct, writing statutes that require licensed professionals rather than systems to issue adverse determinations (Holland & Knight). The law is catching up to what audiences already feel.
Four: judgment is what happens when the data runs out
This is the lesson that changed how I think about my own clinical work.
Preparing a talk about clinical decision-making forced me to define what physicians actually do that a model does not. The answer is not pattern recognition. Models are frequently better at that than I am.
What I do is decide under uncertainty with incomplete information, competing risks, a patient's own values in the room, and a clock. I decide when the data has run out and something still has to happen. At 2 a.m., with a scared family, a history I cannot fully verify, and no bed upstairs, the hard part was never retrieval.
A model narrows the range of reasonable options. It does not choose among them, and it does not carry the consequence.
Say that plainly to a room of clinicians and something shifts. They stop defending their territory and start asking which parts of the job they would happily hand over.
Five: the ending is the argument
TEDx taught me that a talk is not a container for information. The last ninety seconds are the whole thing, and everything before them is setup.
That is true of AI implementation too, and most health systems get the order backwards. They start with the tool and work forward to a use case. The talks that work, and the deployments that work, start with the ending. What should be different for a patient or a clinician when this is done, and then work backward to what is required.
I try to end my AI talks by returning the audience to responsibility. The point is not to make people amazed by the tool. The point is to help them decide what they will do differently when the next tool enters their hospital, clinic, boardroom, or family conversation.
If you cannot state the ending, you do not have a talk. You have slides. The same is true of an AI strategy.
What I tell physicians preparing their first talk on this
Clinicians ask me about this often enough that it deserves its own section.
Pick one idea. Not three. The instinct from grand rounds is to survey the field. A general audience can hold one argument, and they will hold it well. Everything else you know becomes evidence for that one thing or it gets cut.
Cut the credential slide. Your authority comes from the specificity of what you describe, not from a list of positions. One well-chosen clinical detail establishes more credibility than a bio.
Say the uncomfortable thing early. Audiences are braced for a physician to either defend medicine or sell technology. Doing neither in the first two minutes buys you the rest of the talk.
Practice out loud, standing, without slides. If the argument does not survive that, the slides were carrying it, and slides do not carry anything on a stage with a red circle on the floor.
Decide what you want the room to do. Not feel. Do. A talk aimed at a feeling produces applause. A talk aimed at an action produces a follow-up email, and the follow-up email is the entire point.
I still get nervous. Five stages did not fix that, and I have stopped expecting it to. The nerves come from the same place as the care, which is a reasonable trade.
What five stages left me with
A discipline, mostly.
Explain it without jargon or do not deploy it. Attach every number to a person. Answer the accountability question before the accuracy question. Be precise about what judgment is, so you can be precise about what to automate. And know the ending before you build the middle.
That is the speaking discipline behind DR GPT™.
None of that is about artificial intelligence alone. It is about being clear enough to be argued with, which is the only standard that has ever protected patients.
Harvey Castro, MD, MBA is a board-certified emergency physician, 5x TEDx speaker, and author of more than 30 books on AI and healthcare, including AI in Emergency Medicine (Wiley). He serves on Singapore's Ministry of Health Regulatory Advisory Panel and advises the Texas Medical Association's Committee on Health Information Technology.
Book Harvey for your board, leadership retreat, or healthcare AI event: www.HarveyCastroMD.com | /HarveyCastroMD | #DRGPT
Related DR GPT™ reading
More from DR GPT™: who DR GPT™ is, speaking, TEDx healthcare AI, books, media kit, and the healthcare AI board advisory.
Book Harvey Castro, MD, MBA, known as DR GPT™, for healthcare AI keynotes, board briefings, and leadership retreats. Check availability and book.