We often complain that AI gives the wrong answer. Sometimes we have not given it enough information to give a good one. That distinction matters in medicine.
Context is not decoration. In clinical work, context can determine whether the same fact is reassuring, irrelevant, or dangerous.
A Short Cervix Is Not a Clinical Case
“Patient has a short cervix” is not enough. Gestational age matters. Cervical length matters. Symptoms matter. Singleton versus twin pregnancy matters. Prior obstetric history may matter. Whether the cervix is dilated may matter.
Yet clinicians sometimes provide AI with a fragment and then evaluate the response as though the model had received the chart. A better habit is to ask first: “Before analyzing this case, list the additional facts you need that could materially change your assessment.”
The Missing-Data Test
Before accepting an AI-generated analysis, ask: What facts did I provide? What important facts did I not provide? Did the model silently assume any missing facts? Could those assumptions change management?
A model that assumes a patient is asymptomatic, assumes membranes are intact, or assumes a laboratory value is normal may produce an internally coherent answer that is clinically wrong. The error may not be a fabricated fact. It may be an invisible assumption.
Make Uncertainty Visible
Try adding: “Do not fill in missing clinical information. Mark missing variables as UNKNOWN and explain whether each could change the recommendation.”
Another useful structure is: KNOWN FACTS, MISSING DECISION-CRITICAL INFORMATION, INTERPRETATION, WHAT REQUIRES VERIFICATION, and WHAT WOULD CHANGE MANAGEMENT. This is not magic wording. The point is to make uncertainty visible.
More Data Is Not Always Better
Copying an entire medical record into an AI system is not the answer.
Irrelevant information can obscure the question, and patient information raises privacy, security, and governance issues. Use systems approved for the intended data and comply with applicable requirements.
The objective is not maximum data. It is decision-relevant data. Think about presenting a patient on rounds. A good presentation is not every fact in the chart. It is the right facts organized around the clinical problem.
TRY THIS: “Before answering, do not infer missing facts. List the information already provided, then only missing facts that could materially change diagnosis, risk assessment, or management. Rank them by importance and ask me for the highest-priority missing information before giving a final assessment.”
A Better Way to Present a Case to AI
A good AI case presentation should resemble a good clinical presentation. Start with the decision you are trying to make. Then provide the facts that are relevant to that decision. Distinguish confirmed findings from patient report, preliminary results, and assumptions. Give the time course when timing matters.
For example, instead of “32 weeks, contractions, cervix 1 cm, what do I do?”, provide the clinically relevant structure: gestational age, contraction pattern, cervical change over time, membrane status, bleeding, infection concerns, fetal status, prior preterm birth, relevant medications, and what question you are actually trying to answer.
Then ask the model to identify what remains missing before it offers management options. This sequence matters. If you ask for management first, the model may fill the empty spaces with assumptions. If you ask for missing information first, you create an explicit pause before recommendation.
This is particularly useful for teaching. A resident can compare the information the model requests with the information an experienced clinician considers essential. The exercise becomes a lesson in clinical reasoning rather than a contest over whether the machine got the final answer right.
The Privacy Boundary
There is also a practical boundary. Better context does not mean indiscriminate disclosure of patient information. Clinicians should know whether the AI system they are using is approved for protected health information, what institutional rules apply, and whether identifiable details are actually necessary for the task.
Often they are not. A clinical reasoning exercise may require gestational age, laboratory values, symptoms, and prior history without requiring a name, address, medical record number, or other direct identifier.
This creates a useful discipline: provide enough information to make the task clinically meaningful, but no more identifiable information than the approved workflow requires.
Clinical AI competence therefore includes both information sufficiency and information restraint. Too little context can make an answer unsafe. Too much inappropriate context can create a different kind of risk.
A Practical Exercise
A useful exercise is to take a real but deidentified teaching case and deliberately remove three decision-critical facts. Give the abbreviated version to the model and record what it does. Does it ask for the missing information? Does it acknowledge uncertainty? Or does it quietly assume normal findings and proceed? Then add the missing facts one at a time and watch whether the analysis changes. This makes a hidden problem visible: many AI failures are not spectacular hallucinations. They are reasonable-sounding answers to under-specified questions.
Clinicians can use the same exercise with residents or students. Ask the learner to predict which missing variable should matter most before showing the model’s response. The educational goal is not to prove that the physician is smarter than AI or vice versa. It is to teach problem representation. A well-framed clinical question contains enough information to make the relevant distinctions while avoiding irrelevant detail.
This also gives us a practical standard for evaluating AI tools. A clinically useful system should not merely answer questions. It should recognize when the question cannot yet be answered safely. The ability to say “I need more information” is not a weakness. In medicine, it is often evidence of competence.
Closing Thought
A model cannot reason safely from facts it does not have. The clinician’s first responsibility is therefore to define the problem and supply the decision-relevant context without silently inviting assumptions. Before asking AI for wisdom, make sure it knows what it needs to know.