The tasks you should never give ChatGPT

A note before the list. Tools improve, and some of these limits will soften over time. The durable bit is the principle underneath, keep ChatGPT off the jobs where being wrong is expensive and hard to spot. That holds whichever tool you are using.

ChatGPT will have a go at anything. Ask it to do your tax, draft a contract, total a column of figures, diagnose a rash, and it will return something fluent and confident for all four, with no warning that two of them are the kind of thing it should never have touched. That willingness is the whole problem, it does not say no, so you have to know where the line is.

This is not a list of things ChatGPT is bad at to make you feel clever for avoiding it. It is the short list of jobs that play to exactly what it cannot do, where handing them over costs you more time than doing them yourself, or worse, costs you nothing visible until it matters.

The short answer. Keep ChatGPT off any task where a wrong answer is expensive and you would not catch it. It predicts plausible text rather than working things out or looking them up, so the jobs to avoid are the ones that need real calculation, current facts, confidentiality, or knowledge of your actual situation. Everything else, it is genuinely good at.

Why "it can do anything" is the trap

ChatGPT is a confident generalist that guesses. That makes it brilliant at the things where a strong first draft is the goal, and quietly dangerous at the things where there is a single correct answer and it does not know it.

The danger is not that it refuses, it is that it produces a wrong answer with the same calm fluency as a right one, so nothing on the screen tells you which you have got. (If you want the longer treatment of why that happens and how to spot it, we wrote how to tell when ChatGPT is making it up.) So the tasks to keep off it are not random, they are the ones where guessing is either expensive, undetectable, or both.

The tasks to keep off ChatGPT

• Anything where a wrong number matters. It does sums by pattern-matching what a plausible answer looks like, not by calculating, so it gets totals, percentages and multi-step maths wrong in ways that look completely reasonable. A figure that is off by a decimal place reads exactly as confidently as a correct one. For anything that has to add up, use a spreadsheet or a calculator, and if you do ask AI, ask it to show its working so you can check the steps.
• Anything current that you will repeat as fact. It will invent a statistic, a date, or a source to give you a complete-looking answer, because a confident guess scores better with you than an admission of not knowing. For live facts and anything you need to quote, Perplexity is built for it, it finds and cites rather than generates.
• Anything confidential or client-sensitive, on the standard version. On the consumer tier, what you paste can be used to train the model by default and is retained for a window, so client details, contracts, financials and anything under an NDA do not belong in it without real thought. The short version is, if you would not email it to a stranger, do not paste it into a public chatbot.
• Anything that depends on knowing your actual situation. It cannot see your company, your numbers, your relationships, or the politics of the room. Ask it to make a real decision about your specific circumstances and it will fill the gaps it cannot see with confident generalities, which is how you end up with advice that sounds sensible and fits nobody.
• Final-form legal, medical or financial decisions. Drafting a first version, explaining a concept, helping you prepare questions for a professional, all fine. Treating the output as the decision itself is not, since it does not carry the liability, you do, and it has no way of knowing what it does not know about your case.

Anything you genuinely cannot check is the one that catches people out. If a topic is so far outside your knowledge that you could not tell a right answer from a convincing wrong one, you cannot use AI for it safely, because you have no way to catch the error. The tool is only as trustworthy as your ability to verify it.

What it is genuinely brilliant at instead

So you do not come away thinking the thing is useless, because it is the opposite. ChatGPT is excellent for getting unstuck on a blank page, turning messy notes into a structure, reformatting something tedious, explaining a concept you half understand, brainstorming options, and drafting the version you then make yours. Those are the jobs where a fast, confident first pass is exactly what you want, and where a small error costs nothing.

If your main work is writing or close analysis, Claude tends to handle nuance and longer documents better. If it is research you need to cite, Perplexity. If it lives inside Gmail and Docs, Gemini. But for the everyday drafting-and-thinking work, ChatGPT earns its place, as long as you keep it off the list above.

The one question that sorts it

You do not need to memorise the list. You need one question, before you hand anything over: if this is wrong and I do not catch it, what is the damage. High damage and hard to catch, keep it off ChatGPT or check it properly. Low damage and easy to spot, hand it over and move on.

That is the same judgment the Trust Ladder is built on, scaled to where the output is going. Knowing what to keep off the tool is not a limitation, it is the thing that makes everything you do put on it actually worth trusting.

Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it.