The earlier lessons explain what AI does badly. This is the one you actually use: a decision you take before opening the chat, in five seconds, that settles 90 % of cases.
The whole thing in one sentence. It is not the topic that decides, it is
what happens if the answer is wrong. The same question about the same
medicine can be green or red depending on where the answer ends up.
The traffic light
Green — go ahead, no second thought
- Drafting a notice, a poster, an email to a supplier
- Summarising a text you paste in yourself
- Translating something of yours
- Turning text into a table, or the other way round
- Ideas for the pharmacy blog or socials
- Rewriting a leaflet in plain language for you to review afterwards
Red — you do not do this
- Entering anything that identifies a patient
- Asking for a dose, an interaction or a price and using it without the source
- Diagnosing
- Deciding whether someone stops or changes a treatment
- Deciding whether to refer
- Producing a legal or clinical text that gets signed unread
The five questions, when the traffic light does not settle it
For the cases that do not fall clearly into a colour. If a single answer lands on the right, it is not a job for AI:
| Ask yourself | If the answer is… |
|---|---|
| Does this leave my screen? | Yes, unreviewed → no |
| Is there an identifiable person? | Yes → no, no nuances |
| Does it hinge on an exact figure? | Yes → go to the source, not the chat |
| Would I know if the answer were wrong? | No → you need the source first |
| Is it a clinical decision? | Yes → it is yours, not its |
The question most people skip is the fourth. "Would I know if it were
wrong?" If the answer is no, then you cannot use your own judgement as the
filter — and that is exactly when people trust it most, because the explanation
sounds convincing. The less you know about the topic, the more plausible it all looks.
A case that looks green and is red
"Rewrite this leaflet in plain language for an older patient." It sounds like pure drafting, which is the green part of the list. But in simplifying, a model eats the nuances: the administration condition, the renal impairment warning, the "no more than X days". And that text leaves the pharmacy with your credibility behind it.
It is still a good task for AI — with the missing step: reading it against the leaflet before it gets printed. Green to write it, yours to hand it over.
The rule behind the lesson. If the error stays on your screen, use AI
without fear. If the error leaves your screen, AI writes and you check. And
if the error can reach a patient without passing through you, it was not a job for AI.
A full example, and what is wrong with it
The leaflet case above is a green task turning red because of what gets removed when you simplify. This is the other route, and it is the commoner one: a green request turning red because of what the text adds to look good.
What gets asked: "Write me the window poster for this autumn's immune-support campaign. Make it eye-catching."
What it replies
THIS AUTUMN, DON'T STOP
Boost your defences and prevent colds with our vitamin C + zinc pack. Clinically proven to shorten the duration of a cold.
Perfect for the whole family. Now also available with Lemsip for the first symptoms.
–20% this week. Ask your pharmacist.
And now, what is wrong with that reply:
- It is well written, and that is the whole problem. It sounds like a pharmacy poster, it has rhythm, it fits the window and it reads from a distance. Because the request was "write" — textbook green on the traffic light — nobody reads it again before printing. And that is the jump: the traffic light classifies the task you asked for, and what needs classifying is the text that came out. They are not the same, and the gap shows up precisely in the requests that sound most harmless.
- "Prevents colds" is a health claim. Attributing disease prevention to a product has its own rules and is not decided in a shop window. "Boost your defences" is in the same family. The model has invented nothing odd: it wrote what the posters it learned from write, and plenty of those posters could not say it either. An error copied from a thousand examples looks right by majority, which is exactly the trap from the bias lesson.
- "Clinically proven" is an unsourced citation, in your window. It is the same hallucination as lesson 3, only on foam board and facing the street. Nobody is going to ask you for the DOI — but if somebody does, and sooner or later somebody does, you do not have it. And there is one important difference from a chat: a screen can be cleared and a window cannot, and it is signed with the pharmacy's name, which is the one thing you cannot redo.
- And it has named a medicine and put a discount on it. You did not give it the brand name: it put it in because it fits an autumn poster. Advertising a medicine to the public has its own rules — what may be named, on what conditions, with what discounts — and they fall outside this school and this page, but they do not fall outside your responsibility. And "ask your pharmacist" at the end does not fix it: it comes after the promise, the citation and the price.
The fixed poster says almost the same thing and promises nothing: it talks about what the pharmacy does — we review your medicine cabinet, we explain what helps and what does not — instead of what the product does. One line in the request gets you there: "no prevention or cure claims, no cited studies, no medicine brand names".
And the rule that comes out of this, which is the one to take away from Level 1: the traffic light is applied twice. Once to the task, before asking; once to the text, before it leaves the screen. Almost every problem in this school happens between those two moments.
And if your pharmacy is not like that
If you work the shift on your own. There is no second pair of eyes, so the second read has to be yours and at a different moment. It works better than you would think: write it today and read it tomorrow before printing. The distance does almost the same job as another person, because what stops you seeing the error is not lack of knowledge — it is that you have just read the text three times and you are no longer reading it, you are recognising it.
If the person using the AI is not the one who signs. This is normal in a pharmacy with staff: whoever has time writes it and it goes out with the house name on it. So the traffic light has to belong to the pharmacy, not to each person, because if everybody applies their own, the loosest one decides for all. A three-line list taped next to the computer — what may be printed, what gets a second look, what never goes out — is worth more than a whole training session.
If what you write goes on social media. There amber becomes red for a practical reason: there is nobody to take it back. A poster comes down; a post gets shared, screenshotted and forwarded round neighbourhood groups within hours. And the text that performs best on social media is exactly the one that promises most, so the pressure runs the wrong way. Rule that works: on social media you talk about what you do, not about what a product does.
If it is for internal team use. Protocols, counter scripts, the rota, a summary of something new. Here the traffic light is gentler and it is worth saying so, because if everything is red nobody uses it: internal material can be corrected next week and reaches no patient. The one precaution is to date it and record who reviewed it, because today's draft is next year's protocol and by then nobody will remember it came out of an unchecked chat.
If the text is going to a doctor or to an accountant. One thing changes and it changes a lot: it is read by somebody who does know the subject. An approximate paragraph a patient would accept is spotted instantly there — and what you lose is not a poster, it is the pharmacy's credibility on the next phone call. Paradoxically it is the easiest situation to solve: these are short texts with one specific fact in them, and you have that fact in the record or in your software. Let the AI write it; you supply the fact.
When it does not work first time
With this traffic light almost everything comes out amber.
Then you are classifying correctly: amber is the majority colour and it is not an uncomfortable zone, it is the working zone. What needs solving is not "how do I turn this green" but what the source is for that particular amber — the product database, the official gazette, the leaflet, your own protocol. Once the source is clear, checking takes thirty seconds and amber stops feeling like a chore. It is amber with no identified source that paralyses you.
My colleague uses it for everything and nothing has happened.
Probably true, and it does not prove what it looks like it proves. Almost all errors of this kind never show up: a poster with a claim it cannot make hangs there for months, a summary missing the renal-impairment warning gives no signal, and the patient who followed weak advice does not come back to tell you. The absence of visible consequences is the normal state, not evidence that the method works.
I have asked a thousand times and it has never slipped a claim like that in.
It depends a lot on how you ask, and that is the clue: "eye-catching", "make it sell", "make it punchy" are instructions that push straight towards the claim, because in the texts it learned from, striking and promising are the same thing. Change the adjective in the request — "clear", "sober", "informative" — and you will see the poster come out clean most of the time. Half of this problem is solved in the asking.
So do I ask it for the poster or not?
Ask it, of course. Writing posters is one of the things it does best and it saves you half a morning. What changes is that it comes off the printer after passing through you, with three things flagged: no prevention or cure claims, no cited studies, no medicine names. That review takes twenty seconds. The first-week mistake is not using it for the window: it is printing the first thing that comes out because it was well written.
In an emergency I have no time to apply any traffic light.
And you should not: in an emergency the AI does not come into it, and that is the decision, taken in advance and without thinking. What you prepare beforehand are the tools that do work under pressure — the poisons information number, the protocol on the wall, the product database bookmarked — because in two panicked minutes nobody is going to evaluate a reply. If you catch yourself opening a chat with somebody waiting in front of you, that is the signal that the question was not one for a chat.
Before moving on
- I know it is not the topic that decides, it is where the answer ends up.
- I can place a task in green, amber or red without much thought.
- I ask myself the fourth question: would I know if it were wrong?
- I know that simplifying a clinical text eats the nuances.
- No clinical decision comes out of a chat.