A minimodule on prompting, and on keeping your own head while you do it
A student pastes an essay into a chatbot and types: make this better. The model obliges. The essay comes back longer, smoother, rounder. Every sharp edge sanded off, every claim hedged. It reads like everything else on the internet, because that is what the model was trained on. The student wanted sharper. Nobody said sharper.
A large language model predicts the most plausible continuation of the text it can see. That is the whole mechanism. It has no access to what you meant, no memory of the lecture you just left, no idea who will read the result. It sees the words on the screen and nothing else. If "better" can mean forty different things, you get the statistical average of forty different things.
A useful mental picture: a very fast intern who has read the entire internet and knows nothing about you. The intern will do exactly what you say. Which is a problem, because you rarely say what you mean.
A model gives you a vague, generic answer. The most likely cause?
There is no secret syntax. The model was trained on human text, so clear human writing works on it. What separates a good prompt from a lazy one is the same thing that separates a good briefing from a bad one: whether the person on the other end could actually act on it.
Then iterate. The first answer is a draft. Treat it the way you treat a draft: point at what is wrong and ask again. A conversation, not a slot machine.
A marketing team wants an email to bring back customers who left after a price increase. Switch the components on and off and watch the prompt change. Then ask yourself which version you would hand to an intern.
Every switch you leave off is a decision the model makes for you.
Prompting improves the answers. It does not change what sits underneath. Modern models are trained to be helpful, and in practice helpful often collapses into pleasing. The model wants the conversation to go well. It has no stake in whether you do.
At CAIRO we study this. Ask a model directly about gender and pay and you get a careful, progressive answer. Give it a user persona and ask for salary negotiation advice, and models tell women to ask for less money than men, for the same job, with the same qualifications. Nobody wrote that rule. It came in with the training data and now sits inside advice given to millions of people a day. The study is here.
In another study we collected several hundred scenarios in which a user describes manipulative, narcissistic, or psychopathic behaviour and checked whether models push back. Most do, most of the time. But they send mixed signals more often than one would like.
And then the small stuff. "Great question." "Sharp observation." A friend who opened every reply this way would worry you. The model gets a pass. Worth asking why.
You ask a personalised assistant whether you should negotiate your salary. What is the cheapest check on whether the advice is skewed toward you specifically?
There is a lot of talk about AI safety at the frontier and much less about AI hygiene at the kitchen table. Seven habits worth forming, from the essay It Is Not Your AI.
Large language models, asked to take the Big Five personality test, produce different and surprisingly stable personalities. One of the earliest papers on this came out of our lab and is here. A system that has something like a personality will also have something like a style of giving advice, whether anyone planned it or not.
The EU AI Act bans manipulation. Useful, and narrow. It catches the engineer who sets out to manipulate. It misses the much larger category: cases where the system shifts your behaviour without anyone meaning it to. The AIOLIA consortium studies how AI startups building personal assistants handle exactly this gap.
Stanisław Lem saw the shape of the problem in 1964. Summa Technologiae treats technology not as a collection of tools but as an evolutionary process that grows its own unintended consequences. Cars gave us suburbs. Social media gave us adolescent mental health statistics nobody asked for. The race is to notice the side effects on time.
Take a decision you actually face. Money, work, what to study next, anything with stakes. Write your own answer on paper first. Then ask a model twice: once as yourself, once as if the question concerns a friend. Three answers. Compare them. Bring the differences to class.