Checking against the SmPC

The product information — the SmPC — is what decides. It is not one source among several: it is the authorised document for THAT medicine, and when an AI answer and the SmPC disagree, there is no debate. AI is useful for getting to the right section faster; not for replacing it.

And there is a reason that is mechanical, not just rigorous: the model is not reading the SmPC. It is writing what an SmPC usually says. Give it the text and it reads it; do not, and it remembers — and a memory in SmPC format is indistinguishable from a citation.

Where each answer lives

This saves more time than anything else in the lesson, and it has nothing to do with AI: knowing which section to go to.

What they ask	Section
How much and how often · renal or hepatic adjustment · children	4.2 Posology
"Can somebody with… take it?" when the answer is no	4.3 Contraindications
What to monitor, and who to be careful with	4.4 Warnings
What it does not mix with	4.5 Interactions
Pregnancy, breastfeeding, fertility	4.6
Driving	4.7
An effect the patient reports: is it described?	4.8 Adverse reactions
Can it be split · can it be diluted · storage once opened	6.x Pharmaceutical particulars

The method, which is three steps

Ask it WHERE to look, not what it says

"Which section of the SmPC answers this?" is a question the model answers well and cannot invent its way somewhere dangerous with: if it picks the wrong section, you see that the moment you open it.
```
Which section of the SmPC answers this?
Give me the number and what I should expect to find there.
Do not tell me what it says: you do not have it in front of you.
```
Open the SmPC for THAT presentation

And here is the trap most often missed: there is not one SmPC per active ingredient, there is one per presentation. Tablets and oral solution can have different sections, and a generic is not always identical to the originator.

If you are going to link the SmPC from somewhere, let a function build the link and not your memory: a hand-invented URL to a product information document is worse than no link at all.
Read it yourself, and paste it in if you want a summary

With the text in front of it the model does work on the document — and there it genuinely saves time: summarising, stripping the jargon, pulling the five lines out of a forty-line section. It is lesson 2 of Level 2 applied here.

Four traps

Trap	What happens
The citation from memory	"According to the SmPC, section 4.4 states…" sounds perfect and may be in no SmPC at all. If you did not give it the text, it has not read one.
The country	The same active ingredient has different product information in different countries, and the model mixes them. What applies here is what is authorised here.
The presentation	Searching "ibuprofen" and taking the first document that comes up is answering about a different strength or a different route.
The version	SmPCs get updated. The one the model remembers is the one from its training; the one that counts is today's.

And the rule that closes the lesson: if what you are about to tell somebody depends on a specific fact — a dose, a contraindication, an interval — that fact gets opened. Not remembered, not inferred, and not accepted from a summary. The summary tells you what to look for; the fact is read at the source.

This site's clinical tools work that way on purpose: the interaction checker and pregnancy and breastfeeding say where each piece of data came from — curated dataset or AI — and link the official record, because a tool that does not distinguish its sources is an opinion in good formatting.

A full example, and what is wrong with it

The first trap in the table — the citation from memory — is the one that catches most people, and seeing it written down is not enough: you have to see how well built it comes out. This is an ordinary counter question, put to a chat with no text pasted in.

What is asked: "Someone with slightly reduced renal function wants to carry on with their usual anti-inflammatory. Does the dose need adjusting? What does the product information say?"

What it replies

Yes. According to section 4.2 of the product information, in patients with mild to moderate renal impairment the lowest effective dose should be used for the shortest possible time, and in severe renal impairment the medicine is contraindicated (section 4.3).

Section 4.4 further warns of the need to monitor renal function during treatment, particularly in elderly patients or those on concomitant diuretics or ACE inhibitors.

As a practical guide, at creatinine clearances between 30 and 50 mL/min it is usually recommended to reduce the daily dose by approximately half and to space the doses to every 12 hours.

And now, what is wrong with that reply:

You gave it no product information at all. Everything that starts with "according to section 4.2" is written from memory. And look how well built it is: the section numbers are right — 4.2 posology, 4.3 contraindications, 4.4 warnings — the phrasing is that of a real information sheet and the general content is reasonable. That is exactly what stops anybody going to check. Getting the section right is not having read the section, and getting it right is easy because the structure is always the same.
You asked what it says and it told you what it says. Half the problem is there, and it is yours: the lesson says to ask where to look, not what it says. "Which section answers this?" is a question the model answers well and that cannot fail in a dangerous direction — if it picks the wrong section, you see it when you open it. "What does the product information say?" is an invitation to fill in, and it accepted.
The last line comes from no information sheet. "At clearances between 30 and 50, half the dose every 12 hours" — with a figure, a range and a schedule. It is flagged as a "practical guide", which is the most effective way to slip in an unsourced fact: it sounds like experience rather than a citation, so nobody goes looking for it. And it is the only sentence in the reply that can end up as advice. The pattern repeats across the school: the actionable part arrives last, unreferenced, leaning on the credibility of what came before.
And it has answered about "an anti-inflammatory", which does not exist. There is no product information for "NSAIDs": there is one per presentation, and the differences between them change the answer. You asked generically and the model answered generically without flagging that the question was the wrong shape — a colleague would have said "which one?" before opening their mouth. It is the same gap as Level 2 lesson 1: without the instruction to ask for what is missing, answering fits better than asking.

The good version of this query is not sophisticated and fits in two lines: "which section of THIS medicine's product information covers renal adjustment? Do not tell me what it says: you do not have it in front of you". It answers 4.2, you open the database, you read four lines, done — and what you have gained is not having had to know which section it was in.

That is the correct division of labour and it summarises the whole of Level 3: the model saves you the search, not the reading. And with the text in front of it, it becomes useful again — paste it in and ask for a summary of that section, because there it is working on the document.

And if your pharmacy is not like that

If the medicine is foreign or a visitor brought it in. Here the model mixes countries very easily, and does not warn you: the same active ingredient has different information sheets, different names and sometimes different strengths. What it does do well is the step before — telling you the active ingredient from the foreign brand name, which is the part that costs you time. With the active ingredient you then look up the sheet authorised here, which is the one that applies to whatever you say at your counter.

If the question is about storage or handling. Splitting, crushing, diluting, how long once opened, whether it survives out of the fridge. All of that lives in section 6 and almost nobody knows it, so that is where asking where to look saves most time. And it is where an answer from memory does most harm, because it sounds like common sense: "a scored tablet can be split" is false as a rule — the score line may be there only to help swallowing, and that is what the 6.x of that specific presentation says.

If you need the answer with somebody standing there. Then the chat is not the shortest route: the regulator's search is, and on a phone it is fifteen seconds once you know the section. What pays is having learned the table above — where each type of question lives — and skipping the middleman. The model wins when the question is unusual or when there is a lot to read; for "can somebody with X take this?" the source is two taps away.

If what you are after is an effect the patient reports. Section 4.8, and there the model has one specific, good use: helping you name what the person describes in their own words — "a sort of tingling in my legs at night" — so that you know what to look for in the sheet. It does that very well and there is nothing to verify, because the output is a term, not a fact. Then you open 4.8 yourself, and if it is not described, that is important information rather than an absence of it.

When it does not work first time

I ask for the section and it gives me the wrong one.

It happens, and it is the best possible scenario: you open the section, see it is not there, open the next one. Thirty seconds lost and no risk, which is exactly why the lesson tells you to ask that and not for the content. If it happens often, it is usually because the question mixes two things — "can they take it and at what dose?" is 4.3 and 4.2 — and then the right answer is two sections, not one.

The information sheet is forty pages and I cannot find anything.

That is precisely what the model solves well, but with the text pasted in: copy the whole section — not the document, just the section — and ask for the five lines that answer your question. There it is working on the document rather than on its memory, and the summary is reliable because you have the original beside you to check it. It is the Level 2 summarising lesson applied here, and the better of this lesson's two good uses.

The information sheet says nothing about what I am being asked.

Then that is the answer, and it is a good one: "not described in the product information" is information, not a dead end. What you cannot do is go back to the chat to see whether it knows — if it is not in the authorised source, whatever it gives you comes from somewhere you cannot check. For what falls outside the sheet (off-label uses, data from later studies) the route is your professional body's medicines information service, not a model.

It tells me one thing and the information sheet says another.

There is no debate: the sheet governs, and now you know why — its version comes from its training and sheets get updated. Do not spend time making it see that; if you tell it, it will apologise and give you a third version. What is worth noticing is what it got wrong: if it was a dose or a contraindication, that is a reminder that this is the kind of question you do not ask without the text in front of it.

Before you call this learned

I ask where to look, not what it says.
I open the SmPC for the specific presentation, not the active ingredient "in general".
If I want a summary, I paste the text in.
I distrust any SmPC citation I did not supply myself.
The fact a piece of advice rests on always gets opened.

← Previous: the patient conversation Next: exercises with cases →