When AI Explains Something, Keep the Details That Do the Teaching

Suppose you ask an AI assistant to explain why bread dough rises. You get an answer about yeast producing carbon dioxide, with a colourful detour about the history of bakeries. Later, you can picture the old bakery described in the answer. When you try to explain what made the dough expand, the steps are less clear. You remembered a piece of the explanation, but it was not the piece you needed.

The useful question is whether each detail helps the reader do something they could not do before. If the goal is to explain why dough rises, a diagram of gas bubbles in the dough earns its place. A bakery anecdote might be right for a history lesson or a newsletter about food culture. In a short lesson about the mechanism, it can send attention in a different direction. The same detail can be helpful or distracting depending on the job.

Why an interesting detail can get in the way

Educational psychologists use the term seductive details for material that attracts interest but does not help with the learning objective. The detail can be related to the general topic and still be irrelevant to the task. That is what makes it harder to spot than a random interruption: a reader can reasonably assume that a story placed inside an explanation is there to help them understand it.

In a study published in the 2019 issue of Applied Cognitive Psychology, Alexander Eitel, Lisa Bender and Andreas Renkl compared learning materials without these details with materials containing them. Some participants who received the extra details were told that they were irrelevant to the learning goal. The group that received the details without that guidance did worse than the group without the details; the researchers did not find the same disadvantage for the informed group in that experiment.

In a 2021 study led by Bender, participants worked through a longer digital chemistry unit. The researchers again found poorer learning outcomes when the extra details were present and their irrelevance had not been explained. These are experiments on instructional materials. They did not test an AI chat or establish that a label will always protect learning. They do give us a useful editorial question: does the way a detail is presented tell the reader what work it is meant to do?

Keep the example that carries the explanation

Imagine you run a small studio and are writing a guide for new freelancers about spotting a scheduling conflict. You ask an AI assistant for a relatable example. It invents a designer juggling two client deadlines, then gives the designer a favourite café and a story about a cancelled train. The two deadlines belong in the example. The train belongs only if the lesson involves allowing time for travel or handling a delayed handover. The café can go unless it helps explain something the reader needs to decide.

Cutting every concrete detail would make the example worse. Keep the times that overlap and the fact that both deadlines require the same person. Add a second case in which two jobs overlap on paper but one can proceed without that person. The contrast shows what makes the first case a conflict. The reader has a rule they can apply to their own diary, rather than just a story about a hectic day.

One way to edit is to ask what a reader could misunderstand if you removed a sentence. If the answer is ‘they might apply the rule to a case where it does not hold’, the sentence may be doing important work. If the answer is ‘they would miss an amusing aside’, decide whether you are teaching the idea, inviting curiosity or trying to do both. A clearly labelled optional note can give readers a choice about when to follow the detour.

Try it on your next AI explanation

Before asking for an answer, write a finish line in one sentence: ‘After reading this, I want to be able to explain why the dough expands’, or ‘I want to identify a real clash in next week’s schedule’. If the explanation is for somebody else, use the task they need to do rather than the topic you want to cover. Then review the response against that finish line.

• Keep the mechanism. Look for the steps or conditions that explain why the outcome occurs. Ask for any missing step before asking for a more engaging rewrite.
• Keep a revealing example. Choose a case that helps the reader apply the idea. Where useful, add a contrasting case that makes an exception or boundary visible.
• Separate optional context. If a related story is useful for curiosity or motivation, place it after the main explanation and say what it is there for. Leave it out when the reader needs a quick instruction.
• Check what the reader can do. Use a new example that was not in the explanation. If it is for your own study, try it after closing the answer, as in our guide to checking what you remember after studying with AI.

If a reader still gets the new example wrong, look at what they missed. Removing an anecdote will not fix a missing definition or a faulty analogy. Sometimes you need to add a detail: a diagram that shows where the gas goes, or a second example that distinguishes a shared deadline from a shared bottleneck. Aim for an explanation in which the things that catch the eye are also the things that help the reader understand.

Sources

• Eitel, Bender and Renkl, experiment on perceived relevance and seductive details, Applied Cognitive Psychology, 2019.
• Bender, Renkl and Eitel, study of seductive details in longer learning sessions, Journal of Computer Assisted Learning, 2021.