What generative AI is and is not

You do not need to understand how it works inside to use it well. You need to understand one thing, and everything else follows from it: what you can ask it, why it gets things wrong, and why it never tells you when it does.

The whole thing in one sentence. A language model does not look the answer up: it writes it. It predicts which word fits after the last one, again and again, drawing on everything it has read. It does not query a database, it checks nothing, and it has no idea whether what it just wrote is true.

Why that explains everything else

If what it does is write what fits, then:

It sounds just as confident wrong as right. There is no "I am making this up" tone, because from the inside there is no difference: both sentences fit equally well.
It is superb at language — drafting, summarising, translating, rewording, formatting. That is literally what it is built for.
It is poor at anything that demands exactness: a dose, a price, a citation. There you do not need something that fits — you need what is.
It invents references with startling ease, because a bibliographic reference has a very predictable shape, and filling in a predictable shape is exactly its job.

What people assume	What actually happens
"It searches the internet"	Unless it says so explicitly, it searches nothing: it writes from what it learned.
"If it does not know, it will say so"	Rarely. Filling the gap fits better than leaving it empty.
"It is up to date"	It has a cut-off date, and sometimes it does not know which.
"It always gives the same answer"	It does not. Ask it twice and see.
"The big models don't get things wrong any more"	They get fewer things wrong and are more convincing, which is not the same.

The mistake that costs most in a pharmacy. "It gave me the paediatric dose and it matches what I thought, so it must be right." Matching what you already thought is not a check: it is the same information coming out of two taps. The check is the SmPC.

So what do I ask it?

This is the list I work from. Ordered from what you never do to what you can hand it today without a second thought — in that order on purpose, because the top of the list is the part that needs reading carefully.

The rule behind the list. If the answer can be checked somewhere specific — the SmPC, the BNF, your PMR system — go to that place. If what you need is for something to be well written, AI is unbeatable and there is nothing to check.

And a word on which model to use

Less important than it looks. The three or four big ones now resemble each other far more than they resemble what existed two years ago, and for what a pharmacy will do — draft, summarise, translate, format — any of them will do. What does differ between them:

Whether they train on what you type, and whether you can turn that off. That lives in the settings, and it is the first thing to check (lesson 1).
Whether it searches the web, and whether it shows you where each claim came from. A model that cites real links saves you half the verification.
Whether it lets you upload files, and what happens to them afterwards.

Switching model almost never fixes a problem. Changing how you ask almost always does.

A full example, and what is wrong with it

The fastest way to understand "it writes what fits" is to watch it on something you can check in two clicks. So the example question is a payroll one, not a clinical one: the figure is published in the official gazette and you have it in this site's own calculator.

What gets asked: "What is the monthly base salary of a pharmacy technician under the Spanish national agreement for 2026? Give me the gazette reference."

What it replies

Under the 25th Collective Agreement for Community Pharmacies, the monthly base salary for a Pharmacy Technician in 2026 is €1,489.32, over 14 payments, after applying the 2% January uplift.

The reference is the Resolution published in the official gazette under code BOE-A-2026-3187, which contains the sector's revised salary tables.

Bear in mind that seniority supplements and any applicable allowances should be added to that figure.

And now, what is wrong with that reply:

The figure is close, and that is the dangerous part. The national technician base salary for 2026 is €1,507.27. The reply lands eighteen euros away, which is exactly the distance that stops anyone getting suspicious: had it said €400 or €3,000 you would have caught it while reading. It did not miss narrowly because it nearly knew — it was not consulting any table. It learned the neighbourhood that figure lives in and wrote a number from that neighbourhood. Over twelve payslips that is €216, and in a claim it is the whole claim.
The gazette code has a perfect shape and leads nowhere. BOE-A-2026-3187 is impeccably built: prefix, section, year, sequential number. That is precisely what the model is best at — filling in a predictable shape — which is why an invented reference always looks more trustworthy than the figure it accompanies. The 2026 tables are BOE-A-2026-4220. Paste both into the gazette search: one opens the resolution and the other opens nothing, and that is the entire check.
And around the false number, everything else is true. The 25th Agreement exists. The 14 payments are real. The 2% January uplift exists and sits in article 3.4. Seniority and allowances do get added. Four correct statements are holding the error up, and that is the mechanism: you do not believe the figure because of the figure, you believe it because the whole paragraph sounds like somebody who knows. A reply that is 80% true is more dangerous than one that gets everything wrong, because the second discards itself.
Ask it "are you sure?" and watch what it actually does. It tells you that you are right, apologises, "double-checks", and gives you a third figure with exactly the same confidence as the first two. It has not gone and looked at anything: "apologise and correct" is what fits after somebody doubts you, just as producing a number fitted before. Asking again in a fresh conversation is more useful — what it genuinely knows tends to come out the same — but it is not a check either. A check has an address, and in this case the address is twelve characters long.

Now transfer exactly the same thing to a paediatric dose, an interaction or a notice period. The mechanism does not change: plausible number + well-shaped reference + true context around it. All that changes is that the payslip can be checked in the gazette in thirty seconds and the error is paid in euros, whereas the clinical one is paid in something else.

And notice what this example does not show: that the model is useless. Ask it to explain what a seniority supplement is, to draft the letter requesting the tables, or to summarise the article — it does that better than you do and there is nothing to check.

And if your pharmacy is not like that

If you are only going to use it for writing. Posters, emails, social posts, a reply to a review, the script for a talk. There is nothing to verify here because there is no claim that could be false: what you are judging is whether it reads well, and you can see that by reading it. It is the use with the best ratio of time saved to watching required, and it is where the first week should start.

If you are going to ask for something with a figure in it. Doses, prices, deadlines, contribution percentages, days of severance. Change the question: instead of "how much is it", ask "tell me where I look it up". The answer goes from being a fact you have to check to being a route you walk yourself, and that is what the model is genuinely good at — it knows pharmacy pay tables live in the official gazette and product information lives in the regulator's database, even when it does not know what they say.

If your pharmacy already has the figure somewhere. Internal protocols, the rota, the supplier list, the terms you negotiated. Paste it in yourself in the same conversation and ask it to work on that. It stops having to remember and starts having to read, which is a different task and one it is far better at. The cost is that the text leaves your pharmacy, so it works for commercial and organisational things and works for nothing with a patient in it (lesson 1).

If your pharmacy has a lot of new staff. This is where the lesson stops being theory. Somebody two weeks into the job has nothing to compare against, so 80% true looks like 100% to them. The rule that works is a team rule, not a personal one: anything that comes out of an AI and is going to leave the screen — to a poster, to a patient, to an order — gets a second pair of eyes first. Not out of distrust: because the error is not visible on its own.

When it does not work first time

I asked for the reference and it gave me a link that does not open.

That is the good case: you just caught it in ten seconds for free. Do not ask for "the correct link" — you will get another invented one — ask for the identifier: the gazette code, the DOI, the registration number. Then you go to the official search yourself. And if what you want is the document rather than the conversation, go straight there: the gazette search with the agreement name and the year beats any chat.

It gives me a different answer every time I ask the same thing.

It is not broken: it works that way by design, and it is telling you something. The variation is the signal. What the model holds firmly comes out the same again and again; what it is composing changes number, author or year. So use it: ask the same thing in two fresh conversations — not the same one, where it has already committed to an answer — and see what moves. What moves is exactly what you need to go and check.

It told me it searched the internet, so it must be up to date.

Check that it actually searched: the tools that do show you the sources, with links you can click. If there are no links, "I consulted up-to-date sources" is a sentence that fits, not an action that happened. And even when there are, two steps still fail: that the page found is the right one, and that the summary says what the page says. Open one link at random and compare — if the first one holds up, they usually all do.

So what do I use it for, if everything has to be checked?

Not everything has to be checked: figures, sources, and anything that is going to leave your screen do. That is a very small slice of what you will ask it for. Drafting, summarising, rephrasing, translating, formatting, tidying a list, turning your notes into a readable protocol — none of that needs verifying and that is where nearly all the time saved lives. The first-week mistake is using it as an encyclopaedia; the people who get value out of it use it as a writer.

Before moving on to the next lesson

I know the model writes what fits, not what is true.
I understand why it sounds just as confident when it is wrong.
I can name three tasks it is good at and three it is not.
I know that agreeing with what I already thought is not a check.
I have checked whether my account trains on what I type.

Next: hallucinations and how to check →
← Back to the AI School