Five Questions Before You Open the Chatbot

Leave the chat tab alone, just for one minute.

Your hands got there before the rest of you did, bringing a task or a vague feeling that this would go quicker with help.

Most days that's fine, but some tasks deserve the pause. These are the ones where being wrong would cost you, or where other people's interests are involved.

Sam Illingworth writes a newsletter called Slow AI. He teaches what he calls critical AI literacy, which means thinking carefully about how and when you use AI, and he has five questions for that minute. The questions are his, and the notes under each one are mine.

1. What am I giving up by using AI for this? Sometimes nothing, sometimes the practice, and sometimes the conversation you'd have had with a colleague. Name it, then decide whether the trade is fair.
2. Whose view is missing from the answer I'll get? An AI model (the system behind the chat) learns from what people have written down. Most of that is in English, and most of it comes from people with time to write. Ask about a community or a country that's barely in that record, and the answer will sound confident but only tell part of the story.
3. What's the cost if it's wrong? A clumsy sentence in an email costs nothing, but a wrong answer to a medical question or about a legal clause can cost you something real. The higher the cost, the more of the checking you have to do yourself.
4. Who benefits from me using it this way? Sometimes you do, and maybe the person you're writing to does too. Every time, the company that built the tool does. If the only winner is the number that tracks how much you use the tool, notice that.
5. Would I still trust the answer if I knew how it was made? The tool guesses the most likely next words. It tends to agree with you, and it can't tell you what it doesn't know. If knowing that makes you hesitate, then hesitate.

None of the five questions has a right answer. The answers belong to you.

The questions aren't a test you must pass before you're allowed in. They make sure you were the one who decided.

Make AI ask first

Nobody remembers five questions in the moment, so put them in the tool as a checkpoint that AI runs before helping with anything important. In Claude, you can use a Project or a Skill, and in ChatGPT, a Project or custom instructions. These are all places to save instructions that AI follows every time.

Paste this in as the standing instruction.

When I start a message with the word "check", don't answer the task yet. Ask me these five questions one at a time. Wait for my answer before you ask the next one.
1. What are you giving up by using me for this?
2. Whose view might be missing from the answer you'll get?
3. What's the cost if I'm wrong?
4. Who benefits from you doing it this way?
5. Would you still trust the answer knowing how I work?
After the fifth answer, sum up what I said in three lines. Say whether I seem to want to go ahead. Then wait for me to say yes before you do the task. Never answer the questions for me. Never skip one because it seems obvious.

• Save the word "check" for the tasks that matter. "check draft the complaint to the landlord" runs the questions, but "tidy this paragraph" doesn't, and you'll know which is which.
• Answer roughly. One line each is enough, because the questions work by making you give an answer, however rough.
• Let the check change the plan. By question three, the cost of being wrong may look high enough that you'd rather ask a person. Or question one may tell you that you wanted to do this yourself. Either way, the check did its job, so close the tab.

Having the tool ask the questions is a convenience and nothing more.

The tool asks because you told it to, and you could answer "nothing, nobody, none, me, yes" in five seconds and carry on. Mean your answers, or the check does nothing. The check stays short so that you can mean every answer.

Five questions before every important task can also start to feel like a hassle, and a hassle is how people end up giving up on a good tool.

Does the check feel like a ritual? Cut it down to the one question that matters most for you, which for most people is the cost of being wrong.

No setup at all? Five questions on a sticky note beside the screen do most of the work.

Once the check says go ahead, write down what you think before you read what AI thinks.

Credit. The five questions are Sam Illingworth's, from "What Is Critical AI Literacy?" on Slow AI (February 2026). Read it for his full framing.