What an AI assistant can and cannot do, in one page

If you only read one article on this site, make it this one. Most frustration with AI tools comes from asking them to do the wrong kind of task.

Reliable: shaping text you already have

Summarising a long document, rewriting a paragraph in a different tone, turning notes into a list, translating between major languages, drafting a first version of something you will edit. The assistant has the material in front of it, so it rarely invents things.

Usually fine: general explanations

Explaining how a mortgage works, what a word means, how to change a setting on a phone. Check anything that matters against an official source, but for everyday questions the answers are serviceable.

Unreliable: facts it has to recall

Specific dates, statistics, citations, product prices, legal rules in your state, anything that changed recently. Assistants produce fluent, confident, sometimes wrong answers here. Treat these as leads to verify, not conclusions.

Not what it is for: predictions and decisions

An assistant cannot predict the stock market, the weather beyond what a forecast already says, or any random event. It can list considerations for a decision; it should not make the decision.

A simple test

Ask yourself: could the assistant have the information needed to answer this? If the answer is in the text you pasted, trust it more. If the answer would require knowing the future or looking something up that it cannot look up, trust it less.