What should a beginner use AI for first?

The usual advice is to start with something low-risk, which is nearly right and misses the important half. Low risk is not the criterion; judgeability is. A task where you cannot evaluate the output teaches you nothing regardless of how safe it is, because you finish it with an answer and no idea whether the answer was any good. A task where you can evaluate the output instantly gives you a calibration you will use for everything afterwards.

Concretely, that means starting on your own material and your own domain. Summarising a document you have read. Drafting a reply to an enquiry of a type you have handled a hundred times. Reformatting your own data. Explaining something to you that you already understand, so you can see what it gets subtly wrong. Each of these has a known correct answer sitting in your head, which is the fastest verification instrument available and costs nothing to run.

The tasks to avoid at the start are the ones the marketing recommends: research on subjects you do not know, strategy, market analysis, anything where the output is a claim about the world you cannot check. Those are not more dangerous because the model is worse at them. They are more dangerous because you have no way to notice when it is wrong, and a run of plausible unverifiable output builds exactly the wrong kind of trust.

There is one genuinely good first task that is not about judging quality: transformation. Turning a list into a table, notes into an ordered summary, an email thread into a set of decisions and owners. The correctness is largely mechanical and inspectable, the time saving is real and immediate, and it introduces the habit of supplying material rather than describing it — which is the single largest determinant of output quality later.

What to notice while doing this is more important than what gets produced. Notice where it invents a specific. Notice what it does when the material does not contain the answer. Notice how the output changes when you supply an example of what you want versus describing it. Notice how much of the result you had to fix, and whether fixing it took less time than doing it. Those observations are the actual deliverable of the first week.

The last thing to avoid at the beginning is automation. A scheduled or unattended job removes the very thing that makes early use valuable, which is that you see every output. Automate after you can predict what the system will produce, which for most people takes a few weeks of ordinary use and cannot be shortened by reading about it.