Why does figuring things out yourself from scattered sources stick better than just reading someone else's neat summary?

This explores why doing the assembling yourself, rather than receiving a finished summary, seems to produce deeper understanding. The corpus has one direct human study on this, plus several lines of research on language models that suggest why it happens. The direct evidence is large: across seven randomized experiments with more than 10,000 people, those who learned a topic through ChatGPT summaries reported learning less, felt less ownership of what they knew, and wrote advice that independent raters judged sparser and less informative than advice from people who used ordinary web search Does learning from AI summaries produce shallower knowledge than web search?. The surprising part is that the shallowness shows up in what people can produce afterward, not just in how they feel about it.

One likely reason is that summaries lose specificity in a systematic way. General words ("animal") are used more often than specific ones ("beagle"), and language models lean toward frequent phrasings. So every round of AI paraphrasing drifts toward abstraction and wears away the expert-level detail Does word frequency correlate with semantic abstraction?. A summary gives you the general shape and quietly drops the specifics that make knowledge usable. When you read several sources, you meet those specifics again and again, in different forms.

A second reason is that the useful part of a text is often the reasoning behind it, not the conclusion. One line of work describes expert writing as the surface residue of hidden thinking: the self-talk, recall and checking that produced it. Models trained on text with that thinking rebuilt and added back learn reasoning that transfers to new domains Can reconstructing expert thinking improve reasoning transfer?. Research on what models learn from pretraining points the same way. Reasoning ability comes from broad *procedural* knowledge gathered across many varied documents, while recalling facts depends on narrow memorization Does procedural knowledge drive reasoning more than factual retrieval?. Comparing sources makes you see how different people reach and justify their claims. A summary hands you only the claims.

The work of assembling also has value in its own right. Understanding a text means tracking at once what each part says, why it is there, and what matters most. That work is how comprehension is built How do readers track segments, purposes, and salience together?. Reasonable readers also take different things from the same sentence Why do readers interpret the same sentence so differently?, so a single summary settles disagreements you might have learned from. Summaries aren't useless, though. Retrieval systems do better when they first build a global map of a document and then use it to find scattered evidence Can building a document map first improve retrieval over long texts?. That suggests a useful way to read: treat a summary as a map that tells you where to go, not as the place you end up.

A caveat: only the first study is about human learners. The rest is from AI and linguistics research, so read it as a likely explanation, not proof. Still, the pattern is consistent. Understanding seems to come from reconstructing the reasoning and specifics yourself, and summaries tend to remove exactly that.