Using AI Well: Concrete Strategies

There’s a difference between “I want to know something” and “I want to be calmed down.” You can spot reassurance-seeking by three signs: you basically already know the answer; you rephrase the same question several times until the tone of the reply feels right; and the relief afterwards lasts less long than it did last time.

Two small moves reliably interrupt the loop. First: write down your own assessment in one sentence before you send the prompt. It costs ten seconds and restores the right order — your judgement first, the second opinion afterwards. After two weeks you can read back how often you were right on your own. For most people the result is surprisingly good.

Second: cap the number of answers — one question, one answer, then close it, and no more than three follow-ups per topic. Uncertainty doesn’t shrink with more answers; it tends to grow, because every new phrasing brings a new variant into play. Sitting with that briefly is uncomfortable, and it is exactly the point where the loop loses its power.

Chatbots are trained to answer helpfully and warmly. One consequence is that they tend to agree with you — even when your plan has holes in it. Agreement feels like confirmation, but it isn’t. It’s a property of the system, not evidence for your view.

So it’s worth having a simple rule for anything with consequences — money, health, relationships, resigning, confronting someone: before you act on it, you tell at least one human being, or check it against an independent source. Not to hand the decision over, but so that at least one party is involved who has no interest in pleasing you.

Inside the chat, the counter-check helps: “Give me the three strongest arguments against my plan.” That often brings a lot — but it doesn’t replace the person, because the model still follows your framing.

The most intense conversations happen at night: usage peaks sit at the edges of the day, and the longest sessions often fall between two and five in the morning. That’s no coincidence — at night the threshold is lower, the mood heavier, judgement tired, and nobody else is awake.

So the most effective rule is structural rather than a matter of willpower: the device charges outside the bedroom, and an alarm clock wakes you. If your phone spends the night beside you, you have to resist all over again every time; if it charges in the hallway, you don’t have to at all.

And if the urge comes anyway: write the question on a piece of paper next to the bed. What felt urgent at three is often an ordinary to-do at nine — or nothing at all.

Regularly do one thing deliberately without AI: write a text yourself, do your own research, make a difficult phone call without pre-drafted sentences, think a decision through to the end on your own. Once a week is enough.

The reason isn’t principle. Skills you hand over permanently lose their contours — and with them the feeling of being able to do something yourself. That feeling is the best protection there is against dependence on any tool. The friction is inconvenient, and it trains you.

A resolution like this holds best if you tie it to a category rather than a day: “Anything I write to friends, I word myself.”

Chatbots are well suited to anything that involves knowledge, structure or practice: having something explained until it clicks; sorting out your thoughts and structuring texts; rehearsing a difficult conversation beforehand; practising a language; getting informed before you see a professional. In that role they’re a rehearsal space and a tool — and rather good at it.

They are not suited to crises. When it comes to acute thoughts of suicide, panic, violence, or a situation where you no longer feel safe, you need people — immediately, and not as second choice (the numbers are in the chapter “Finding help”). Nor do they replace psychotherapy, trauma work, or a medical assessment about medication.

The third limit is the quietest one: it shouldn’t be your only confidant. Not because that’s forbidden, but because a single channel for everything important loses every counterweight. As a rule of thumb: excellent as a rehearsal space, unsuitable as the last word.

Staying honest here means saying this: there are no tried-and-tested programmes with solid evidence of effectiveness for problematic AI use — the topic is too young for that. The individual building blocks, however, are not new.

Self-monitoring (writing down what actually happens), stimulus control (taking triggers out of your field of view) and questioning automatic thoughts are core parts of what performs consistently well in studies of related digital patterns — problematic internet and gaming use. Transferable doesn’t mean proven, but it is considerably more than a gut feeling.

So don’t expect a transformation. Expect an overview. And if nothing has moved after a few weeks even though you wanted something to change, that isn’t a failure — it’s a good moment to bring someone in.