When NOT to use AI for a task

We build support sites for AI platforms. We use these tools every hour of the working day. And there is a growing set of tasks where our honest advice is: close the tab.

Not because AI is dangerous in a science-fiction way. Because on these specific jobs it is simply the wrong instrument, and using it costs you something.

When you cannot verify the answer

This is the master rule and everything below is a special case of it. If a claim matters and you have no practical way to check it, generated text is a liability. You are not getting an answer; you are getting a fluent, confident guess wearing the costume of an answer.

Ask yourself before you paste: if this is subtly wrong, what happens? If the answer is "someone gets hurt, sued, or misled," stop.

Legal, medical and financial specifics

General education, fine. "What questions should I ask my doctor about this" — genuinely useful. "Is this contract clause enforceable in my jurisdiction" or "what dose is safe" — no. The failure mode is not obvious nonsense; it is a plausible answer that is out of date, jurisdictionally wrong, or missing the exception that applies to you.

When you are supposed to be learning

If the task exists so that you learn something — a course exercise, a language, your first data structure — delegating it removes the entire point. The struggle is the mechanism. You can use a model as a tutor that explains and quizzes you. You cannot use it as a substitute for the reps and expect the skill to arrive.

When it is other people's data

Anything containing someone else's personal information, health details, private messages, or your employer's confidential material deserves a hard pause. They did not consent to being pasted into a third party's system. "But the terms say they do not train on it" is not the same as "the people in this document agreed to this."

When the human relationship is the content

Condolences. Apologies. Performance feedback. A love letter. A reference for someone who trusted you.

The value in these is not the prose quality — it is that you sat down and did it. Outsourcing the sitting-down empties the gesture, and being caught doing so is worse than writing something clumsy. Clumsy and sincere beats polished and hollow every time.

When you need one true fact, fast

For a specific number, date, quote or citation, a search engine and a primary source beats a generative answer. Faster, verifiable, and no risk of a confidently invented reference. Models are for synthesis, not lookup.

When speed is an illusion

Some tasks feel faster with AI and are not. If verifying the output takes longer than doing the work, you have added a step. This is common for short tasks you already know how to do — writing the four-line email yourself is faster than prompting, reading, editing and re-reading.

When you have not decided what you think

If you use a model to form your opinion rather than pressure-test it, you will end up with the consensus view every time, delivered persuasively. That is fine for facts and corrosive for judgement. Decide first, then argue with the machine.

What good judgement looks like

The reliable test is one question: am I using this to do more of my thinking, or less of it?

More is a tool. Less is a habit. The people getting the most out of these systems are, almost without exception, the ones with the clearest sense of where they stop.

Knowing when not to reach for it is a skill. Tick that one off first.