An explanation can seem clear while I am reading it and fall apart when I try to explain it to someone else. Usually the missing part is a relationship: how one event causes another, why a step is necessary, or which assumption makes the conclusion follow.

Rozenblit and Keil studied this gap in people’s accounts of everyday mechanisms. Participants’ estimates of their understanding often fell after they attempted detailed explanations.[^1] Recognition had made the mechanisms feel more familiar than the participants’ explanations could justify. Trying to explain exposed what a first judgment had missed.

Understanding should therefore involve some capacity to work with an idea. I may need to explain its reasoning, apply it to a new case, or recognise where it fails. The appropriate demand depends on the subject. Understanding a poem does not require the same performance as understanding a pump. Both require more than recognising familiar words.

Studies of self-explanation offer evidence about how this capacity develops. Michelene Chi and colleagues observed differences in the way students explained worked physics examples. More successful learners related steps to principles and identified gaps in their comprehension.[^2] In a later study, students prompted to explain a text about the circulatory system developed better understanding than students who read it twice without that prompting.[^3] Explaining can become part of learning, rather than merely a demonstration after learning is finished.

The explanation need not begin elegantly. “I do not see why this follows” is often a useful first sentence. It locates a problem that a smooth summary might hide. I can then return to the source with a more precise purpose.

What I already know affects this process. How People Learn II describes how prior knowledge can support learning and can also interfere when an existing account is inappropriate.[^4] A new claim may fit easily because it is well supported, or because it resembles something I already believe. Those experiences feel similar enough to require further judgment.

Drawing a diagram can help me inspect an account, especially when several relationships must be considered together. The diagram still has to contain relationships I can explain. In one study, Karpicke and Blunt found that retrieval practice produced better learning from science texts than the concept-mapping activity used for comparison, including on questions requiring inference.[^5] The result concerns those learning conditions. It gives us a reason to examine what a learner does with a diagram, rather than treating the finished diagram as proof of understanding.

Writing works in a similar way. A tidy paragraph can conceal borrowed reasoning. Revising it becomes useful when I identify the claims I can defend and the transitions I cannot. Sometimes the sentence that needs attention is the one that reads most easily, because its ease has allowed me to avoid the question it should answer.

AI can supply explanations and outlines quickly. The effect on learning depends on how the help is used. A 2025 survey of knowledge workers by Hao-Ping Lee and colleagues reported relationships between confidence in generative AI and self-reported critical-thinking effort.[^6] As a survey, it does not establish that AI use caused a general decline in people’s ability to think. It does suggest examining where users place their confidence and effort.

A field experiment by Hamsa Bastani and colleagues offers a more specific warning. High-school mathematics students using an unrestricted GPT interface performed better with the tool available but worse on later unassisted performance; a tutor designed with safeguards mitigated that cost.[^7] Completing an assisted exercise and acquiring the ability to complete it independently were different outcomes. The design of the assistance mattered.

I can use an explanation to finish a task or to develop an ability. These purposes may overlap, but I should know which one matters on the occasion. If I am trying to learn, I need chances to attempt the reasoning myself, discover mistakes, and correct them. Assistance that removes every difficulty may also remove the practice I needed.

Understanding is seldom complete. I can explain one part of a subject well and remain uncertain about another. Describing those limits makes the account more useful. It tells me where I can proceed and where I need further work.