Context-Aware AI vs. ChatGPT for Clinical Reasoning

A junior doctor I know types into ChatGPT, on his phone, during a break: "70 year old male, K+ 5.8, peaked T-waves, on furosemide and spironolactone, eGFR 28. What's the first step?"

ChatGPT returns a textbook answer. Calcium gluconate, insulin–dextrose, recheck K+ in an hour. Discontinue spironolactone. Consider salbutamol nebs.

The answer is correct. The answer is also slightly useless.

He had to restate the case

Every value he typed is already in the patient's chart on the ward computer thirty feet away. The act of typing the case into ChatGPT is the act of breaking the link between the AI's answer and the patient. The next time he wants to ask about this patient — say, an hour later when the recheck K+ comes back at 5.4 — he has to re-restate.

The restatement is not just friction. It is a filter, and he is the one applying it. Look at what made it into those twenty-eight words and what did not: no time since the ECG, no baseline potassium, no urine output, no note that the spironolactone was started nine days ago, no digoxin question, no mention of the ACE inhibitor he is also on. He did not omit those to save typing. He omitted them because, in the two seconds available, they did not seem relevant — which is precisely the judgement he was consulting the AI to check.

An AI given a case summary is not reasoning about the patient. It is reasoning about the doctor's model of the patient, and the errors it is most likely to make are the ones already latent in the summary.

There are no citations he can verify

The answer is plausible, internally consistent, and produced without any indication of source. For a hyperkalemia case this is fine. For a case where the standard differs between AIIMS and NICE, it is not fine — the resident would need to know which one ChatGPT defaulted to, and the answer doesn't tell him.

This is why source attribution matters more than confidence scores. A model that says "high confidence" has told you nothing you can act on. A model that says "AIIMS protocol, hyperkalaemia, section 3" has told you where to look in nine seconds, and you either agree with the source or you don't.

The model doesn't know what is normal for this patient

A K+ of 5.8 in a stable CKD patient with chronic mild hyperkalemia is a different clinical situation than a K+ of 5.8 in an acute presentation. ChatGPT cannot tell — because it has not seen the patient's prior K+ values, their baseline, or their trajectory.

This is the single largest gap, and it generalises well beyond potassium. Almost every number a ward doctor acts on is interpreted against that patient's own history rather than against the population range. A creatinine of 1.9 is an emergency in one patient and a Tuesday in another. A systolic of 105 is fine unless this man has run at 150 for a decade. The reference interval printed beside the result is the least informative comparison available, and it is the only one a context-free model has.

And it cannot do anything

There is a fourth gap that is easy to miss because it does not feel like a failure of intelligence. The answer arrives as text, in an app, and stops there.

The calcium gluconate does not get ordered. The recheck at one hour does not get scheduled, so it depends on the doctor remembering across the next sixty minutes of a busy shift. The spironolactone does not get flagged for the pharmacist. The handover at 8 p.m. has no idea any of this happened, so the night resident inherits a potassium of 5.8 with no note that it is being actively managed and a repeat is pending.

Every one of those is a place the plan can quietly fall on the floor, and none of them is fixed by a better answer.

What "context-aware" has to mean to earn the name

The phrase is being applied to a lot of products that do not deserve it. A usable test — five questions, all of which should be yes:

1. Does it see the trend, not the value? Prior potassiums, the slope, the baseline.
2. Does it know what is already prescribed? An AI that recommends stopping spironolactone without knowing it is prescribed is guessing.
3. Does it cite something checkable? A named guideline the doctor can disagree with.
4. Is it scoped to one patient? Ask "what's the plan" and it should be unambiguous which patient you mean, with no restatement.
5. Does the answer land somewhere? An order staged, a task created, a line in the handover — not a message in a chat log that disappears when the app is closed.

A context-aware AI consult fixes the first four by construction. It runs from inside the patient's record, so the doctor doesn't restate. It cites the source guideline it pulled from (AIIMS, ICMR, NICE) so the doctor can verify. And it sees the patient's prior values, so a K+ of 5.8 with a baseline of 5.4 is a different recommendation than a K+ of 5.8 with a baseline of 4.1.

Where the general model still wins

It would be dishonest to end without saying this. For a question with no patient attached — what is the evidence for early beta-blockade in this condition, remind me of the Light's criteria, what does this drug interaction actually do — a general model or a literature tool is genuinely excellent, and pretending otherwise would be selling something.

The distinction is not "good AI versus bad AI." It is that a question about medicine and a question about this man in Bed 7 are different questions, and only one of them can be answered from a chat box thirty feet from the chart.

This is the difference between an AI search engine and an AI workspace. The search engine is useful in the same way Wikipedia is useful — when you already know what to ask. The workspace is useful in the same way a senior at your shoulder is useful — when you don't.