Learning With AI: Let the Worked Example Fade

An AI tutor walks you through a calculation. Each line makes sense, and you can explain the answer while the example is on screen. Then you close the chat, try a similar problem and cannot decide where to start. The explanation helped you follow someone else’s reasoning. You still need practice producing the reasoning yourself.

There is a useful stage between asking for the full answer and attempting a blank page. Ask the AI for another example with one step missing. Fill the gap and explain what it does. If you can do both, try a problem with two steps missing. The idea is to remove enough help that you have a decision to make, while leaving enough structure to see where that decision fits.

What the research tested

In a 2002 paper on the transition from example study to problem solving, Alexander Renkl, Robert Atkinson, Uwe Maier and Richard Staley compared a gradual removal of solution steps with alternating complete examples and problems. Their experiments found benefits on closely related problems. In those settings, removing the final steps before the initial steps was more favourable than removing them in the opposite order. The approach is often called fading.

A 2003 study by Atkinson, Renkl and Mary Margaret Merrill combined fading with questions that asked learners to identify the principle behind a solution step. Across two experiments, the authors reported benefits for both closely related and less similar problems. These studies tested instructional designs, years before today’s chatbots. Using a chat to create and deliver this kind of practice is a practical application of the research; the studies do not establish that an AI tutor will produce the same results.

Start with a full example, then leave a gap

Imagine you are learning how to calculate a print run from the number of usable items you need. In a hypothetical exercise, a studio needs 80 usable cards and assumes that exactly four out of every five printed cards will be usable. The worked example shows the relationship: four fifths of the printed quantity must equal 80. Divide 80 by four fifths to find that the studio would print 100 cards under the exercise’s assumptions. Check the result by taking four fifths of 100. Real production losses can vary; the point here is to practise reasoning backwards from a stated yield.

For the next problem, ask for new numbers with the same structure. If the studio needs 60 usable cards and expects three out of four to be usable, the AI might show the relationship and the division, then leave the result and its check for you. Ask yourself why the number printed must be larger than the number needed. That question can expose a mistaken operation even if you have followed the layout of the first example.

On the next round, leave the calculation and check blank, with only the relationship supplied. Then try one where you write the relationship yourself. If you get stuck, ask for a question about the missing step: ‘Which quantity is the usable fraction taken from?’ That gives you something to think about before the chat supplies an answer. If you still cannot proceed, ask to see one more step and try a fresh problem afterwards.

Check the decision behind the arithmetic

A learner can repeat the right calculation for the wrong reason. If every practice problem asks how many cards to print, dividing may become a habit. Try a related question that gives the printed quantity and asks how many cards will be usable. Now the reasoning runs in the other direction. Explain which number is known, which is unknown and why the operation changes. This is a useful point to introduce the approach in our guide to mixing practice problems so you have to choose a method. Fading changes how much of a solution you supply; mixing changes whether the worksheet tells you which method to use.

The AI’s example needs checking too. Use easy numbers at first and verify the relationship by working backwards from the answer. Ask the chat to keep the story and assumptions consistent between the question, hints and answer. If it changes the yield halfway through or makes an arithmetic mistake, fix the example before drawing conclusions about your own understanding.

Run a short practice session

• Ask for one complete example and a separate practice problem with the final step omitted. Ask the AI to wait for your attempt before revealing that step.
• Explain what the missing step achieves. If your answer is wrong, ask for a hint that points to the specific decision you missed.
• Remove more support only once you can explain the previous gap. If you stall, restore one hint and try a different example with the same structure.
• Finish with an unfamiliar problem away from the chat. Reopen the explanation afterwards to check which decisions you made independently and where you still needed help.

This works as a design question for anyone teaching a process, too. If a colleague can complete a task only while copying your example, decide which decision they should make next and leave that part for them to try. The aim is to make the support useful enough that it can gradually disappear.

Sources

• Renkl, Atkinson, Maier and Staley, From Example Study to Problem Solving (2002)
• Atkinson, Renkl and Merrill, Transitioning From Studying Examples to Solving Problems (2003)