In November 2025, the AMA adopted a policy that quietly redefined what it means to be a competent physician. AI literacy is now mandatory. Not encouraged. Not a nice-to-have elective tucked into an afternoon. Mandatory, for every medical student, every resident, and every practicing physician.
I have spent the last few months talking to program directors about this. The reaction is almost always the same. A nod. Then a pause. Then the real question: "Okay, but what does that actually mean for our program on Monday?"
That gap is the whole problem. So let me close it.
First, AI is not one thing
The most common mistake I see is treating "AI" as a single skill you either have or you don't. It isn't. In medicine it breaks into four very different things, and they show up in your day in different ways.
Machine learning is the risk score quietly running in your EHR. Deep learning is the model flagging a nodule on the chest CT. Natural language processing is the ambient scribe turning your visit into a note. Big data is the population layer underneath all of it.
You do not need to build any of these. But you cannot supervise a tool you cannot name. A physician who lumps all of this into one word called "AI" is going to trust it in the wrong places and distrust it in the wrong places.
The debate about "should we use it" is already over
Here is what nobody at the policy table wants to say out loud: your trainees are already using these tools. Today. On real patients. The question was never whether they would. The only question is whether we teach them to use it well, or let them figure it out alone at 2am on a service with thirty patients.
The AMA policy is really an admission of that reality. Adoption already happened. Curriculum is just catching up.
The framework: three tiers, ten competencies
When you read the mandate closely, it implies a structure. I think of it as three tiers.
Tier 1 is foundational. These are the non-negotiable skills every physician needs, regardless of specialty. What AI actually is. Where your data goes when you paste a note into a chatbot. How to write a prompt that gets a useful answer instead of a confident wrong one. The baseline that keeps you and your patients safe.
Tier 2 is where clinical judgment lives. This is the hard middle. Detecting bias in a model that was trained mostly on one population and quietly fails on another. The male breast cancer case is the one I keep coming back to: a tool that performs beautifully until it meets the patient it never saw in training. Reading a radiology AI output and knowing whether to trust it. Interpretation and explainability. This tier is where good physicians separate from passive users.
Tier 3 is leadership scope. Who is liable when the model is in the loop. How HIPAA actually applies at the query level. What your institution should adopt, what it should refuse, and how to govern the difference. FDA pathways for software as a medical device. Model lifecycle: training cutoffs, hallucinations, what happens when the tool reaches past its own knowledge. Not every physician needs to live here. But every program needs people who can.
Ten competencies across three tiers. That is the shape of what "mandatory" now means.
So what do you actually do Monday
Pick the tier where your program is weakest and start there. Most are strong on enthusiasm and weak on Tier 2, the judgment layer, because that is the part you cannot learn from a vendor demo.
This is exactly why I built MeDucation AI. Not to add another tool to the pile, but to teach the judgment the mandate now requires. The policy finally caught up to what fellows have been quietly doing for two years. The programs that move first will graduate physicians who lead this. The ones that wait will graduate physicians who get managed by it.
This is the first piece in a series breaking down all ten competencies, one at a time, with the real clinical example behind each. Which of the three tiers is your program weakest on right now?
Watch the full episode on YouTube: https://www.youtube.com/watch?v=mmwJfwUZO70
