When the patient brings an AI answer

It happens more and more: somebody arrives at the counter with their phone in hand, or a printed screenshot, saying "I asked the AI and it told me...". This is the lesson that closes the circle on everything you have seen in this level — hallucinations, bias, automation bias — but seen from the other side: you are no longer the one using the tool, you are the one who has to respond to somebody who used it without the judgement you already have. It is no longer a rare event either: it happens often enough now that it deserves its own routine, the same way any other recurring counter situation eventually earns one.

The whole thing in one sentence. This is not about telling them they are wrong, nor about agreeing just to avoid conflict: it is about separating what the answer gets right from what it does not, out loud, with the patient right there, because that is the only thing that builds real trust instead of wearing it down.

Why this situation is different from any other counter query

When somebody asks you something directly, you control the starting point. When somebody brings you an answer already written out, the starting point is theirs, and dismissing it outright — "that's worth nothing, don't listen to that" — usually produces the opposite effect to the one you want: the person does not stop trusting the AI, they stop trusting you, because they feel corrected without being told why.

Common reaction	Why it does not work
"What you read is worthless, AI doesn't know about this."	It is a generalisation that is not even true — the hallucinations lesson already showed AI gets things right most of the time — and it sounds defensive. The patient notices, and from then on listens less, not more.
Agreeing just to avoid an argument, even when the answer is wrong.	The most comfortable option short term and the most dangerous medium term: if the answer had a clinically relevant error, staying quiet to avoid friction lets through exactly what your professional judgement exists to catch.
Explaining the whole technical reason why it gets things wrong.	Real information, but not what the patient needs at that moment: they want to know what to do about their medication, not a class on how a language model works.

What actually works: separate and explain the why, not the how. "This part is right — and I'd actually tell you the same — and this other part does not fit your specific case because…" is a sentence that validates the impulse to look things up without endorsing the part that is wrong, and leaves the patient with better judgement for next time they ask an AI something, not just a corrected answer for today.

Four phrases that work, and why

There is no need to improvise every time: there is a structure that keeps showing up in answers that work well, and it is worth having it ready in advance rather than building it from scratch each time at the counter with the patient waiting.

"It's good that you looked into it before coming in." The opening line that changes the tone of everything that follows the most: it acknowledges the initiative instead of questioning it, so what you say next lands as an addition, not as marking an exam.
"This part matches what I would tell you." Naming the agreement explicitly, when it exists, is what stops the conversation reading as "the pharmacy versus the AI". It is not just courtesy: it is true almost every time, because most general answers about common symptoms are reasonably correct.
"What you read did not account for your [specific detail], and that changes things." Naming the specific detail, not a vague "there are more factors", is what turns the correction into something checkable: the patient can verify on their own that detail — their other medication, their age, a pregnancy — is real and relevant.
"Next time you ask something like this, tell it about [what was missing] too." Closing with this turns a one-off case into a habit the patient carries with them, and it is the only one of the four that stays useful even when you are not there next time they think to ask.

A full example, and what is wrong with it

A patient with reflux arrives with a printed answer about what to take, generated by an AI from a description of their symptoms that they wrote themselves.

What is printed, summarised: "For occasional acid reflux you can try an over-the-counter antacid. If symptoms are frequent — more than twice a week — the usual approach is a proton pump inhibitor for a few weeks. Avoid heavy meals before bed and raise the head of the bed."

What is behind it

The patient, in fact, has been on clopidogrel for two months following a recent cardiovascular event — a detail he did not mention to the model because it did not seem relevant to a question about his stomach. The answer he brought is correct in general — the content about reflux is reasonable and states nothing false — but incomplete for him specifically: there is a documented interaction between some proton pump inhibitors and clopidogrel that reduces its antiplatelet effect, and it is exactly the kind of detail that only surfaces if it is asked about explicitly, not when reflux is asked about in general.

And now, what is wrong with the two easiest reactions:

Dismissing the whole answer would be a mistake, because it is broadly right. There is nothing wrong with the content about reflux on its own; the problem is not what it says, it is what it could not have known because nobody told it. Telling the patient "the AI got it wrong" would be inaccurate and would teach him the wrong lesson.
Accepting it without checking would be worse. This is exactly the case where professional judgement adds something no generic answer can: knowing what this specific person takes, which is information that lives in the pharmacy, not in a chat.
What actually works is saying both things in the right order. First, that the answer is generally well written and there is nothing odd about it. Second, that there is a detail of his the model did not have — the clopidogrel — and that detail changes the specific recommendation, not the general information.
And along the way, something reusable gets taught. Not "don't use AI", but "when you ask it something about your health, tell it about all your medication, not just the immediate symptom" — which is exactly the same idea the anatomy-of-a-prompt lesson explains about context, applied to somebody who has not taken this course.

And if your pharmacy is not like that

If it is a chronic patient you already know well. You have the advantage of already knowing their history without having to ask from scratch, so the case in the example is easier to resolve: you can cross-check the answer against what you know in seconds. The risk here is not lacking the data, it is assuming you have already cross-checked it just because you know them, without actually doing it.

If it is somebody new, passing through, that you do not know at all. Here the risk is the opposite: you have no history to compare against, so the only defence is actively asking what else they take before commenting on what they printed out — exactly what the patient in the example did not tell the model, and will not tell you either unless you ask.

If the patient arrives already defensive, expecting to be told they are wrong. Always start with what the answer gets right, even if it is little. Validating first lowers the guard and makes what comes next — the correction — actually get heard, instead of sounding like you are competing with what they read.

If it is an answer about something outside your remit — a diagnosis, not a treatment. Here the right move is not correcting the medical content, which is not your call, but redirecting clearly: "what you're asking is a diagnosis, and your doctor needs to assess that; what I can look at with you is whether what you already take fits with what you read."

If it is an answer about a product you yourselves sell. Watch out for the opposite bias to the usual one: the temptation to accept the part that recommends something you have on the shelf, without applying the same scrutiny as to the rest of the answer. The standard does not change because the outcome suits you.

When it does not work first time

The patient gets defensive as soon as I say something does not fit their case.

Try changing the order: instead of starting with what is missing, always start with what it gets right, and explicitly name that you read the answer carefully, not that you are dismissing it out of hand. Defensiveness almost always comes from feeling their judgement is being questioned for having asked, not from the content of the correction itself.

I do not have time at the counter to read the whole printed answer.

You do not need to read it all: ask first what else they take besides what they came in to resolve, and only read the part of the answer that touches that area. Ninety percent of the risk in these cases lies in the interaction with what they already take, not in the rest of the content.

I get the feeling they trust what they read more than what I tell them.

This is the automation-bias lesson seen from the outside: the text sounds confident, and your spoken answer, if you hesitate while saying it, sounds less confident even when it is more correct. Do not hesitate when correcting a specific detail you know for certain — confidence in tone, here, is a legitimate tool when what you are saying is actually verified.

The answer they brought is completely fine, there is nothing to correct.

Say so explicitly, with the same ease you would use to correct a mistake: "what you read is right, I'd tell you the same thing." Reinforcing when it gets things right is part of building the patient's judgement for next time, not just flagging when it fails.

Should I tell them not to use AI for health questions?

That is advice that does not hold up and that they will not follow anyway: people already do this and will keep doing it. What actually helps is teaching them what to tell the AI so the answer is more useful — all their medication, not just the symptom — and to always bring that answer to the pharmacy before acting on it, which is exactly the behaviour this very case is reinforcing.

What if it is a family member asking on behalf of somebody else, not the patient themselves?

The same method applies, with one nuance: check more carefully whether the family member really has full information — complete medication list, allergies — because it is common for precision to be missing when asking on somebody else's behalf. If you are not sure about a detail, say so explicitly before commenting, instead of filling the gap yourself with a reasonable-sounding assumption.

Before moving on

I separate what the answer gets right from what it does not, always in that order.
I never dismiss a whole answer just for coming from AI, nor accept it whole for the same reason.
I actively ask what else the person takes before commenting on what they brought.
I use the moment to teach what to tell the AI next time, not just to fix today's answer.

← Revisit: ChatGPT, Gemini or Claude
Next: when NOT to use AI, the risk traffic light →
← Back to the AI School