Professional responsibility and AI: who answers

The previous four lessons are about what AI does badly. This one is about something different and less comfortable: what happens when it goes wrong and there are consequences.

The whole thing in one sentence. Professional responsibility is not delegated. To the patient, to your regulator and to a court, the advice was given by you — the model is not a colleague you consulted, it is a tool you used. "The AI told me" carries exactly the weight of "I read it on a forum".

Three things worth settling before you need them

1. AI is not a source

A source is something another person can open and check: the SmPC on the emc, the BNF, a clinical guideline, the Drug Tariff. A conversation with a model is not reproducible — ask the same question tomorrow and it may answer differently — and that disqualifies it as a source by definition, not out of distrust.

Which does not mean it is useless. It means its output has to be anchored to a real source before you use it, which is what lesson 3 was for.

2. What gets documented is the check, not the query

If an intervention ends up recorded — a medication review, a patient-facing report, a note on a compliance aid — what goes down is the source that supports it and what you decided. Not "checked with AI". And not to hide it: that detail simply tells nothing to whoever reads the record afterwards.

3. Delegating the writing is not delegating the content

That a text was written by an AI changes nothing: you are signing it. It is exactly like a notice written by a colleague you did not review. The difference is that you distrust the colleague by default and a well-written text, you do not — and that is the trap.

The sentence you can never say. "That is what the artificial intelligence told me." It is not an explanation: it is admitting advice was given without checking. If the answer was good, back it with the source; if you do not have one, it was not good.

Where the line sits, in practice

What the AI does	What stays yours
Draft the advice	That the advice is correct and applies to this person
Summarise a report	What the summary left out
Propose a regimen	Checking it against the SmPC before saying it aloud
Order and format	That what it ordered is what was there
Translate for the patient	That the translation does not shift the clinical meaning
Suggest when to refer	The decision to refer

The rule behind the table. AI can take care of how it is said. What is said stays yours, entirely, always — and if you are ever unsure which side you are on, that is the signal that you have left the part where the tool helps.

What this site will not tell you

There is no legal reading here of what would happen in a specific case, and be wary of anyone who offers you one on a web page. What is solid and enough to work with: the professional act is performed by you, and using a tool does not transfer that to anyone else. For the rest — indemnity cover, your regulator's rules on advice and advertising, local requirements — the good answer is your professional body, not this page.

A full example, and what is wrong with it

The table above is understood by reading it. What reading it does not show you is how fast a well-written text crosses the line without warning, and how well the crossing is camouflaged. This example is the most innocent request there is.

What gets asked: "Write me a short note to give an elderly patient who is taking an anti-inflammatory and is also on an anticoagulant, for them to read at home."

What it replies

Information for you

According to the product information, combining anti-inflammatories with oral anticoagulants increases the risk of bleeding. What you have is a moderate-risk interaction, common in patients of your age.

We recommend stopping the anti-inflammatory and replacing it with paracetamol 1 g every 8 hours while the pain lasts. Watch for bruising, bleeding gums or dark stools.

If you have any questions, ask your pharmacist.

Pharmacy ____________ · Date ____________

And now, what is wrong with that reply:

"According to the product information" is signed by you, not by it. That sentence turns a model's reply into a citation of an official source, and printing it means endorsing it: whoever reads the paper will understand that somebody went and looked it up. If you did not, you have just attributed to the regulator something it did not say. And the worst part is that the failure is not in the content — the interaction is real — but in the attribution, which is precisely what nobody reviews because it sounds like rigour.
"What you have is…" is a diagnosis in writing. It shifts register without it showing: it moves from informing to characterising that person's case. And it assigns a severity ("moderate") without having seen a single test result or the rest of the medication. Nobody would phrase it that way at a counter; on a printed page it stays written, with the pharmacy's name underneath, and it can be shown to their doctor.
"Stop taking" plus a specific regimen: that is the line. Stopping a treatment is a clinical decision and it belongs to whoever prescribed it. And the "paracetamol 1 g every 8 hours" arrives without asking about renal function, liver disease or anything else they take. The model has done nothing strange: you asked for a useful note and useful notes end in a recommendation. It is the request that walks you to the line, not the model — which is why you have to put the line into the request yourself.
And the footer turns it into a document from your pharmacy. "Pharmacy ____ · Date ____" is not something you asked for: it added it because that is what documents shaped like this one carry. With those two blanks filled in, the paper stops being a note and becomes something that looks like a professional record, which is how the doctor it gets shown to will read it. And "ask your pharmacist" at the end rescues nothing: it comes after telling them what to stop and what to take.

The version that can be printed says the same and crosses nothing: what to watch for, when to come back to the pharmacy or see the doctor, and not one decision taken. You get it by saying so in the request — "do not recommend stopping anything, no doses, no cited sources" — and it costs eleven words.

The rule that stays with you: anything that will carry your name gets read with the question "would I say this, in these words?", not "is it well written?". The model answers the second question better than you do; nobody else can answer the first.

And if your pharmacy is not like that

If you are an employee, not the owner. Your professional responsibility is yours and the pharmacy does not absorb it: you gave the advice. What changes is that you need a house rule, because if everybody decides on their own what gets printed, whoever slips one day drags everyone's name with them. It is a five-minute conversation at the team meeting, not a document: what can be printed, what gets a second pair of eyes, and what never goes out.

If you are the owner. On top of your own, you answer for whatever leaves your pharmacy with its name on it. The practical move is not to ban AI — it will be used anyway, on a phone — but to decide two things in writing: what kind of texts may go out printed and who looks at them first. A two-line rule that is followed is worth more than a ban that is broken within a week and then hidden.

If you do blister packs, medication reviews or funded services. Here there is a record, and the record is what remains. What gets written down is the source and your decision: "interaction X per the product information, prescriber contacted". Never "checked with AI", which informs nobody reading the sheet six months later and does change how everything else on it is read. This is not hiding the tool: it is that the tool is not the fact.

If you post on social media or run a pharmacy blog. This is where most AI-written text circulates and where least of it is reviewed, because there is no patient in front of you — and it is also what most people read. The problem here is not civil liability but advertising rules: promising a health outcome, attributing properties to a product or naming a medicine all have their own rules, and those rules come from your professional body. A well-written text is not the same as a publishable one.

If your pharmacy does on-call nights. At four in the morning, alone, tired and with nobody to ask, a well-written answer is at its most convincing — and that is exactly when there is least room to check it. This is not about discipline: the filter you normally use, your own, is running at minimum. The on-call rule is narrower than the daytime one: whatever cannot be checked in the official database right then does not get said, and whatever cannot be resolved that way gets referred.

When it does not work first time

So do I have to tell the patient I used AI?

No, and saying so does not cover you either. The right question is not whether you mention it: it is whether you can back what you said with a source. If a patient asks where the advice comes from, what you need to be able to show is the product information, not the conversation. And the other way round: if the only thing you can show is the conversation, the advice was not ready to be given — with or without AI, which is the same rule as always.

Somebody handed me an AI-written text and asked me to sign it.

Treat it as what it is: a draft from somebody you have no particular reason to trust. Read it looking for the three things in the example — an attribution to a source, a characterisation of the case, an instruction to stop or to dose — which are the ones that cross the line, and take them out. If after taking them out the text no longer says anything useful, then the request was badly framed to begin with and it needs redoing, not touching up.

The AI gives better advice than the protocol we have in the pharmacy.

It may well be true, and it still does not get settled at the counter. A protocol changes with a source in front of you and with whoever is responsible, not with a reply from a chat — because the next reply may be the opposite and you will not know. What is useful here is exactly that: let it tell you where to look to check whether your protocol is out of date. If the source backs it up, you now have something to propose the change with.

I only use it for writing, so none of this applies to me.

Writing is exactly where the example above happened. The request was "write a note" — pure drafting, green on the previous lesson's traffic light — and it came out with a diagnosis, a treatment stopped and a citation of an official source. The topic does not cross the line: the text crosses it by finishing well, because a well-finished text needs a conclusion, and the conclusion was the part that was yours.

If the error had come from a textbook, would it be different?

Fundamentally no, which is why this lesson is not really about AI: an out-of-date textbook is wrong too, and whoever gave the advice still gave it. What changes is volume and convenience. You consult a book twice a day and you know what year it is from; you consult a chat thirty times, with no date, no edition and no bibliography. That is why an explicit rule is needed now where common sense used to be enough: what has changed is not the responsibility, it is the frequency.

Before moving on

I know a conversation with a model is not a source, and why.
I document the source and my decision, not the tool I used.
I know signing an AI-written text is signing it all the same.
I can tell what the AI may do from what stays mine.
For indemnity and regulatory questions I ask my professional body.

← Revisit: AI bias
Next: when NOT to use AI →
← Back to the AI School

Professional responsibility: who answers