Learning With AI: Make a First Guess Before the Lesson

You ask an AI assistant to explain why a room with bare walls sounds different from one with curtains. It gives you a polished answer. You read it, nod and move on. Now change the order. Before you read, give yourself a moment to predict what would happen to a handclap in each room. Would one sound louder, last longer, or simply sound different? The explanation now has a question to answer and a first guess you can revisit.

That small change is worth trying when an AI assistant is your route into an unfamiliar subject. A first guess makes your starting point visible. A later check shows whether the explanation changed it. The exercise works best as a private, low-stakes attempt with reliable feedback close at hand. There is little reason to spend five minutes inventing a confident answer when a tentative sentence will do.

What pretesting studies found

In a 2009 study by Lindsey Richland, Nate Kornell and Liche Sean Kao, participants read an essay about vision. Some attempted questions about concepts before reading; comparison participants were given more time to study. Across five experiments, the pretested items that participants initially missed were answered better on the later test in the pretesting condition. The researchers also tried directing attention to the concepts in the comparison conditions, including giving people the questions without asking for answers in one experiment.

A separate 2009 study by Kornell, Matthew Jensen Hays and Robert Bjork examined unsuccessful attempts with made-up general knowledge questions and word associations. In those tasks, trying before seeing the answer helped later performance compared with reading the question and answer together. The correct answer followed the attempt. There is a difference between that setup and leaving someone to wrestle indefinitely with an unfamiliar problem.

Neither study tested conversational AI. The findings justify trying the order of events: attempt, study and check. They do not tell us that a chatbot will write a good prequestion, that every guess improves learning, or that a result on the tested items implies a better understanding of the whole subject. Those distinctions matter when you turn a narrow finding into a habit you can use.

Choose a question you can actually resolve

The handclap example gives you something to look for in an explanation. “Explain sound” gives you no comparable point of reference. Before you begin, pick one distinction, prediction or cause. Imagine you run an online shop and want to learn how spreadsheet filters work. A good prequestion might be: if you hide some rows with a filter, what happens to the total in a particular formula? Name the formula or show a harmless practice sheet so the answer is checkable. Then look up the function in the documentation for your own spreadsheet application or test it on a small sheet.

If your lessons begin with a chatbot, give it the source you plan to use. Ask for one question that the source can answer and for enough context to make a reasonable attempt. Tell it to wait while you write your guess. You can skip the tool altogether at this stage if you already have a question. The aim is to make contact with the material before the explanation supplies its own framing.

Make the correction explicit

After studying the answer, put your first thought beside the source. Suppose your prediction was “a filter always changes the total”. The useful correction is more specific: identify which calculation you were using, how it treats the filtered rows, and what you would check the next time. A vague note such as “I got it wrong” leaves too much work for your future self.

If the assistant gives a plausible answer but cannot show where the source supports it, leave the question open. A quick demonstration, original manual or human instructor can settle some gaps. Other questions need extra context before anyone can answer them. You can end a practice session with a well-specified uncertainty; there is no benefit in turning a tentative guess into an equally tentative correction.

Check what remains after you close the chat

It is easy to test the whole exercise on one short lesson. Keep the record light enough that you would use it twice.

• Write one sentence answering the question before you read. Mark it as a guess, or write down what you would need to know to answer.
• Read the relevant passage and record the exact difference between your guess and what the source supports.
• Later, answer the original question without the chat. Try a fresh example too if you need to use the idea in your work. Keep those two results separate: remembering the answer and applying it are different tests.

Our article on putting a recall check before the next lesson deals with a later moment in the learning process. A prequestion comes before instruction and gives you a problem to listen for. A recall question comes after instruction and shows what you can bring back. You can use both without making a short lesson feel like an examination.

Choose a subject where a first guess is cheap and the correction is easy to find. If you still need the AI explanation beside you to answer the fresh example, you have learnt something about what to practise next. That is a more useful outcome than a chat full of answers that seemed clear at the time.

Sources

• Richland, Kornell and Kao (2009): The pretesting effect.
• Kornell, Hays and Bjork (2009): Unsuccessful retrieval attempts and subsequent learning.