When not to use AI

Most writing about AI is about getting more out of it. This one is about the opposite, because knowing when to close the tab is a skill, and almost nobody teaches it.

It is also the part of the course people tell me they were most relieved to be given permission for. There is a quiet assumption in the air that if you are not using AI for everything, you are doing it wrong. You are not.

When the stakes are high and the verification is hard

AI is confidently wrong in the same tone it is confidently right. That is manageable when you can check the answer in a minute. It is dangerous when checking is slow, expensive, or beyond you.

So the test is not "is this important". It is "if this were wrong, how long would it take me to notice". A first draft of a newsletter fails safely; you read it and you see the problem. A medication dose, a legal deadline, a structural calculation or a tax position fails quietly, and by the time it surfaces the damage is done.

Use it for the newsletter. Do not use it for the dose.

When the information should not leave your control

Anything you type into a general AI tool leaves your building. Whether it is retained, who can see it, and whether it trains a future model all depend on the provider, your plan and your settings, and those change.

That is a reason for a rule rather than a case-by-case judgement, because case-by-case judgement fails when you are tired and in a hurry. The rule that works: if you would not put it in an email to a stranger, do not put it in a chatbot.

Names, dates of birth, client files, medical detail, anything under a confidentiality obligation. If your workplace has its own rules about this, those come first, and they exist for a reason.

You can still use AI on that work. You just describe the shape of the problem instead of pasting the specifics. How to stop AI training on your data covers the settings side.

When the point of the task is that you did it

Some things are not outputs. They are the doing.

A condolence note. A speech at your daughter's wedding. A letter to someone you have hurt. An apology. The recipient is not grading your prose, they are receiving evidence that you sat down and thought about them. AI can produce a better-written version of all of these, and it will be worse, because the thing that made it worth anything was the effort you removed.

This is not sentimentality. It is a straightforward reading of what the task is for.

When you are using it to avoid thinking

There is a difference between "I know what I want to say and I want help saying it faster" and "I do not know what I think, so I will see what the machine says and agree with it".

The first is delegation. The second is outsourcing your judgement to something with no stake in the outcome and no memory of you.

The tell is usually your own reaction. If you read the output and think "yes, that is it, tidier", you were delegating. If you read it and think "I suppose so", stop. You have not formed the view yet, and the draft is now standing in the way of you forming it.

When it is genuinely faster to do it yourself

Prompting well takes a minute. Reading and fixing takes longer. For a two-line reply to a client you already know, you will finish before you have finished explaining what you want.

People new to this often lose time for the first fortnight because they route everything through AI, including the things that were never slow. Route the repetitive and the blank-page tasks. Leave the quick ones alone.

When you cannot check it and it will carry your name

If you would not be comfortable being asked, in public, where a claim came from, do not publish it. That applies to statistics, to legal or medical detail, to anything about a named person, and to anything you are about to send to a client under your own signature.

This is the same standard you would apply to a keen new staff member who reads quickly and makes things up when unsure. You would not stop using them. You would check their work before it went out.

What this adds up to

A short list you can actually hold in your head:

• High stakes and slow to verify: no.
• Confidential or identifying detail: not pasted in.
• The effort was the point: no.
• You have not formed a view yet: think first, then draft.
• Faster by hand: do it by hand.
• Going out under your name: check it before it does.

Everything else is fair game, and that is most of the week.

Knowing where the edges are is what makes the middle usable. People who never find the edges either use AI for nothing, or use it for something they should not and get burned once badly enough that they stop entirely. Both of those are worse outcomes than a clear, boring list.

Common questions

Is it wrong to use AI for personal messages?

For most of them, no. For the ones where the effort was the point, a condolence note, a wedding speech, an apology, yes. The recipient is not grading the prose. They are receiving evidence that you sat down and thought about them, and that is the part AI removes.

What should I never type into a chatbot?

A rule beats case-by-case judgement here, because judgement fails when you are tired. If you would not put it in an email to a stranger, do not put it in a chatbot. Names, dates of birth, client files, medical detail, anything under a confidentiality obligation. If your workplace has its own rules, those come first.

How do I know if a task is too high-stakes for AI?

Ask how long it would take you to notice if the answer were wrong. A newsletter draft fails safely because you read it and see the problem. A medication dose, a legal deadline or a tax position fails quietly, and by the time it surfaces the damage is done.

Is it faster to use AI for everything?

No, and people new to it often lose time for the first fortnight by routing everything through it. Prompting takes a minute and fixing the output takes longer, so for a two-line reply you will finish before you have finished explaining what you wanted. Route the repetitive and the blank-page tasks, and leave the quick ones alone.