What AI is good at, and what it is not

A language model is a very good first-draft writer and a very poor last word. It has read an enormous amount of text and learned what usually comes next. That makes it excellent at shape and unreliable about truth.

Hand it these: rewriting something you already wrote, turning rough notes into clean prose, summarising a long document you will still skim, drafting a reply you will edit, listing options you had not thought of, explaining a concept at whatever level you ask for.

Keep these yourself: any fact you would be embarrassed to get wrong, anything about your own company's numbers, anything a client will sign, and any decision that is actually yours to make.

The failure you will meet first is confident invention. Ask for the phone number of a government office in Lahore and you may get a real street with a wrong number, written in exactly the same calm tone as everything else. Nothing in the output marks the difference. That is the whole problem in one sentence: the tone is identical whether it knows or not.

The second failure is blandness. Ask a vague question and you get a vague answer that reads like a brochure. Most people conclude the tool is weak. It is not. The input was.

So the working rule for this course is simple. Use it where a wrong answer costs you two minutes of editing, not where a wrong answer costs you a client. Inside that boundary it will save you several hours a week, starting this week.