Learning With AI: Make a First Attempt Before You Read the Answer

A freelance designer asks an AI assistant how to calculate the percentage of enquiries that turn into bookings. The answer is clear. She follows it in a spreadsheet and gets the figure she needs. A week later, she opens a new sheet and hesitates: should the denominator be all enquiries or only the ones she replied to? She had completed the task, but the explanation had never asked her to make that choice herself.

There is nothing wrong with looking up a formula when the job is to finish a report. If the goal is to remember how to choose a denominator next time, however, the same interaction can be arranged differently. Make a small attempt before seeing the answer. Then use the assistant to examine that attempt, and check later whether you can apply what you learnt.

What the memory experiments tell us

In their 1978 experiments on the generation effect, Norman Slamecka and Peter Graf compared reading words with generating them from cues. Participants in the generation conditions performed better on the memory tests reported in those experiments. The word ‘generation’ here refers to what the learner produced. It has nothing to do with the computer producing text on the learner’s behalf.

The effect is also more specific than a general claim that effort improves memory. A 1991 study by James Nairne, Glenn Riegler and Michael Serra examined memory for items and memory for their order. In their experiments, generation could help memory for the items while impairing retention of their sequence. That distinction matters when you need to remember a procedure. Being able to name the steps is a different test from putting them in the right order.

Neither paper tested present-day AI assistants, spreadsheets or freelance work. The practical suggestion is an inference: if you want a learner to produce something later, make room for them to practise producing it now. Keep the task narrow enough to check, and let the result of that check guide what you do next.

Ask for feedback on a choice you made

Return to the designer’s report. Before opening the chat, she could write one sentence: ‘To measure bookings from all enquiries received in October, I would divide October enquiries that eventually booked by all October enquiries.’ That sentence makes an assumption visible. It groups enquiries by when they arrived and measures whether they eventually booked. If the business question is instead about bookings made during October, the method needs a different definition.

She can then ask the assistant to find an example where her definition behaves differently from another plausible one. An enquiry arrives on 30 October and turns into a booking on 3 November. Which report counts it, and when? Working through that one case does more for the exercise than asking for a longer explanation of percentages. The learner must apply the rule and can compare the result with the intended measure.

A first attempt should be easy to revise. Write a rough answer, list one assumption and ask for a counterexample. If you cannot get started, request a clue or a worked example. You can then try a similar problem with the example closed. The point is to spend effort on the decision you want to learn, without turning a missing prerequisite into a guessing game.

Check what you can do when the chat is closed

The next day, try a related case with different details. The design studio could compare enquiries from two months when several bookings are still pending. Ask whether the two groups have had the same time to convert. If you remember the arithmetic but miss the difference in follow-up time, you have found a specific gap to practise. An immediate sense that the explanation was clear would not have located it.

If the task has an essential sequence, test the sequence. For example, give a colleague three unordered steps: define the group, count qualifying bookings and compute the rate. Ask them to arrange the steps and explain what goes wrong if they count first. If they will use a checklist every time, let them use that checklist in a second exercise. Memory and competent performance with a reference are both useful; the one to measure depends on the job.

• Name the skill. Decide what you want to be able to explain or do later without reopening the original conversation.
• Attempt one decision. Write a prediction, a calculation or a next step before seeing the full answer.
• Get specific feedback. Ask which assumption matters and check the answer against a reliable example or source.
• Try a changed case later. See whether you can apply the idea when the numbers, wording or sequence differ.

Our guide to checking AI meeting notes against the original decision applies a related discipline: keep a trace of what you thought before the polished version arrived. For learning, that trace gives you something concrete to correct. A good AI interaction can finish the report and leave you better prepared for the next one, if it gives you a chance to make the choice yourself.

Use that extra step selectively. A one-off formatting fix may be quicker to look up again. A judgement you make every week is a better candidate for practice. Choose one recurring task and compare what you can explain a day later after reading an answer with what you can explain after trying first.

Sources

• Slamecka and Graf, The generation effect: delineation of a phenomenon (1978).
• Nairne, Riegler and Serra, Dissociative effects of generation on item and order retention (1991).