There are two very different ways for a physician to use artificial intelligence.
The first is:
“Tell me what to do.”
The second is:
“Teach me what I don’t know. Show me what I am missing. Challenge what I think I know.”
Those approaches may look similar on a computer screen.
Cognitively, they are almost opposites.
NVIDIA CEO Jensen Huang recently explained how he uses AI in an interview with Fareed Zakaria. Huang said he does not ask AI to think for him. He asks it to teach him things he does not know or help him solve problems he could not otherwise reasonably solve. He described good questioning as a highly cognitive skill and said that when AI gives him an answer, he does not simply accept it. He challenges it and sometimes gives the answer to another AI for criticism.
For clinicians, this distinction is enormously important.
It may define the difference between AI dependence and Clinical AI Competence.
The Physician Who Asks Better Questions
Physicians have always been professional questioners.
What is the diagnosis?
What else could this be?
What am I missing?
Does this study actually apply to my patient?
Why is the fetal heart rate changing?
What would make me deliver now rather than tomorrow?
What evidence supports this recommendation?
What would change my mind?
A good Maternal-Fetal Medicine consultation is often less about immediately knowing the answer than knowing which questions must be answered before an answer is defensible.
That is why Huang’s observation matters.
Prompting AI is sometimes treated as a technical trick. Learn the right formula, insert the right words, and the machine gives you a better answer.
That is too shallow.
The more important skill is problem formulation.
A weak prompt might be:
“What should I do with this patient?”
A stronger clinician asks:
“What information is missing that could materially change management?”
Then:
“What are the strongest competing interpretations?”
Then:
“Which guideline recommendations apply, and which claims must I verify in the original source?”
And finally:
“What evidence would make your conclusion wrong?”
Now AI is not replacing clinical reasoning.
It is being used to provoke more of it.
Do Not Use AI Only for Things You Already Know How to Do
Here I would modify Huang’s advice slightly for medicine.
He argues against using AI as a crutch for tasks you can already perform.
There is wisdom in that. But physicians can also use AI productively on familiar tasks when it provides a second cognitive pathway.
I can review a manuscript.
I have done it for decades.
But AI may notice an inconsistent denominator buried in Table 3. It may compare the abstract with the results and find that the conclusion overstates the primary outcome. It may identify a reference that does not support the sentence attached to it.
I can interpret an obstetric paper.
But I can ask AI:
“Find the three strongest reasons my interpretation could be wrong.”
That is not outsourcing expertise.
It is using AI to stress-test expertise.
One AI Answer Is Not Enough
Huang described another habit that physicians should recognize immediately.
He compares answers from different AI systems, much as someone might seek multiple medical opinions.
That can be useful, but medicine requires an important qualification.
Three AI models agreeing does not make something true.
Models may share training data, common assumptions, similar published literature, and similar failure modes. Their errors are not necessarily independent.
So I would not ask three models a question and take a majority vote.
I would do something more demanding.
Ask AI #1:
“Analyze this problem and identify your conclusion.”
Give that answer to AI #2:
“Critique this analysis. Identify major factual errors, unsupported assumptions, missing evidence, and alternative interpretations.”
Then ask AI #3:
“Compare these two analyses. Do not choose a winner by consensus. Identify the claims on which they disagree and tell me what primary evidence would resolve each disagreement.”
Then go to the evidence.
That last step is essential.
Multiple AI opinions can identify where uncertainty lies.
They cannot vote reality into existence.
Make AI Attack Its Own Answer
There is an even easier method that requires only one model.
After receiving an answer, do not ask:
“Are you sure?”
That often produces a more elaborate version of the same answer.
Instead ask:
“Assume your answer is wrong. Make the strongest evidence-based case against it.”
Then:
“What evidence would distinguish between your original answer and this alternative?”
That is a much more useful interaction.
For an ObGyn, imagine using this when reviewing a difficult fetal heart tracing, counseling about timing of delivery, examining a controversial paper, or evaluating a proposed quality-improvement policy.
The purpose is not to let AI make the final decision.
The purpose is to expose your assumptions and its assumptions before the decision is made.
AI Can Make a Good Clinician More Curious
This is where I think the fear that AI inevitably makes physicians intellectually lazy becomes too simplistic.
AI certainly can be used lazily.
Ask for the diagnosis.
Accept the answer.
Copy the assessment.
Move on.
That is cognitive outsourcing.
But imagine another physician.
She asks AI to explain a statistical method she has never fully understood.
She asks it to compare three competing explanations for an unexpected result.
She asks it what important question she forgot to ask.
She gives it a paper and asks it to identify where the data do not support the authors’ conclusion.
She challenges its answer.
She opens the original sources.
She asks another model to attack the first model’s reasoning.
And then she decides.
That physician is not thinking less.
She may be thinking about more things than she could reasonably have examined alone.
Whether AI improves or degrades thinking therefore depends partly on how we use it. Huang’s claim that his own cognitive skills are improving is his personal assessment, not evidence that AI use generally improves cognition. The effects of AI on learning and critical thinking remain an active research question.
But his workflow points toward a valuable principle.
The New Skill Is Not Prompt Engineering
Prompt engineering matters. Specificity, context, clear objectives, examples, and iteration can improve AI output.
But I increasingly think the phrase is too narrow for medicine.
A clinician could become excellent at writing prompts and still be dangerously incompetent with AI.
The larger skill is Clinical AI Competence.
That means knowing:
what to ask;
what information the model needs;
what it may be missing;
how to challenge its answer;
what requires independent verification;
which sources deserve trust;
when another model might help;
and when the physician must reject what the machine says.
The best AI user will not necessarily be the physician with the cleverest prompt.
It may be the physician who asks the best second question.
Try This Tomorrow
Take one real clinical, research, or teaching problem that is difficult enough to deserve thought.
Do not ask AI:
“What is the answer?”
Try this sequence instead:
1. Teach me what I need to understand to analyze this problem well.
2. What am I missing that could materially change the answer?
3. Give me your best analysis and clearly separate facts, inferences, and recommendations.
4. Now assume your analysis is wrong. Give me the strongest evidence-based critique of it.
5. What claims in either answer require independent verification?
6. What primary evidence would best resolve the remaining disagreement?
Then verify the important claims yourself.
That is not using AI as a crutch.
That is using AI as an intellectual sparring partner.
The Real Risk Is Passive AI
Medicine does not need physicians who can obtain AI answers.
Soon everyone will be able to do that.
We need physicians who know what to do after the answer appears.
Question it.
Challenge it.
Compare it.
Verify it.
Ask what is missing.
Ask what would make it wrong.
Then apply clinical judgment.
Huang’s most important lesson for medicine may therefore be much larger than prompting.
Do not measure AI competence by how quickly you can get an answer.
Measure it by how intelligently you interrogate the answer once you have it.
Because in medicine, the goal was never simply to have answers.
It was to know which answers deserve to be trusted.
Clinical AI Competence #12
Clinical AI Competence is an ongoing ObGyn Intelligence series about using AI effectively, critically, and safely in clinical medicine, research, education, and patient care.
Subscribe to ObGyn Intelligence for the next Clinical AI Competence lesson.
Source for Huang quotations: CNN, Fareed Zakaria GPS, interview transcript, July 13, 2025.