"AI saves time" is a claim. You can test it by doing one task twice: once by hand (you, plus a calculator or a spreadsheet, but no AI) and once with AI. Then compare what each way cost. Most people count only the seconds the AI took to answer. That is the smallest part.
Before you start, write a prediction: how long will each way take, and which will score higher? Predicting first turns the activity from "see what happens" into a test, because you can see where your guess was off.
About even on time. (Within 10% of your by-hand time counts as even.)
Try making the check shorter or longer. When the answer is easy to check, AI can save time. When checking means redoing the work, it often can't. Quality is the fourth number, and this calculator doesn't show it: it comes from your score against the key.
An invented student ran a data task both ways. By hand: 16 minutes, 5 of 6 answers right. With AI: 2 minutes writing the prompt, 1 minute waiting, 9 minutes checking, 3 minutes fixing.
When you do the same task twice, the second try is easier, because you've already seen the data and you know the answers. If you do the hand version first, the AI version gets a head start: you can check it quickly because you already know what the right answer looks like. That makes the AI look better than it would for someone starting cold. You can't remove this effect from one person, but you can notice it: if some teams do AI first and some do by hand first, compare the class times. When your own timing is unfair, say so in your conclusion.
Real data is messy. Four problems turn up again and again:
Faced with a trap, a tool may fix it, drop the row, leave it in, or point it out. Which one it does varies by tool, and the same tool can do different things on different tries. It often doesn't tell you what it did. That makes the trap dangerous: your answer comes back looking clean, and you can't tell whether it was cleaned the way you'd have done it.
Two habits help. First, ask how many rows went in and how many came out, and compare with the file. Second, before you ask for any analysis, ask the AI to find problems without fixing them, so you decide what to do:
Label every row. Rules for this survey: no trip is longer than 90 minutes, and the same student shouldn't appear twice.
A chart can look professional and still be wrong, because it was built from the uncleaned file. Check each bar against a count you did yourself. Check that the bars start at zero, since a bar chart that starts higher exaggerates small differences. Chart titles and axis labels usually come from your column headers, so don't be surprised when a title matches a column name. Look at the headers in the file before you decide the AI made something up.
Using a calculator, a map or a calendar is offloading. So is asking an AI to do part of a task. Offloading isn't good or bad. It is a trade: you spend less effort now, and you give up something. What you give up depends on what the task was for.
| Ask yourself | Why it matters |
|---|---|
| Can I check it faster than I could do it? | If checking is quick (look up five cells in a grid), handing off can pay. If checking means redoing the work, the AI saved you nothing. |
| What happens if it's wrong and I miss it? | A wrong lunch count costs you a bad chart. A wrong dose, grade or deadline costs far more. The higher the cost, the more of the checking must be yours. |
| Am I supposed to be learning this? | If the point is to get better at it, doing it with the AI means you skipped the practice. Do it first, then use the AI to check or explain. |
A quick test for the third question: after the AI has helped, cover the answer and try to explain or redo it yourself. If you can't, you got the result but not the skill. That is fine for a task you'll never need to do again. It isn't fine for something you'll be tested on, or something you'll be expected to do without the AI.
Five invented students (not the ones in your class puzzle) need one weekly study slot, and each one filled in a grid of free and busy times. You paste the grid into an AI and ask for the slots when all five are free. It says: "Monday 4:30 works for everyone."
Compare that with the lunch survey. There, checking an answer means recomputing it from the file, which can take almost as long as doing it yourself. The same AI can be a good deal for one task and a bad deal for another. Cost of checking is a property of the task, not of the tool.