한국어

Using AI to Summarize Readings: Do You Learn Less?

At a glance
  • People who got the same facts as an AI summary reported learning less deeply than people who read the web pages themselves.
  • They spent about a third less time reading, and the advice they wrote afterwards was shorter and more generic.
  • AI summaries often leave out key points, and when a summary gets something wrong, that can become what you remember.
  • Use the summary as a map before you read, and afterwards close everything and recall it in your own words.

AI summaries save time, but if a summary is all you read, you learn less deeply. Across seven experiments with more than ten thousand people, those who learned from an AI summary felt they had learned less than those who read web pages themselves, and the advice they wrote afterwards was shorter and more generic. This held even when both groups got exactly the same facts. Use the summary as a map before you read, and afterwards close everything and recall what you learned in your own words.

What changes when you study from an AI summary?

People who got a summary spent less time with the material and felt they had learned less. That held even when the content was identical.

Melumad and Yun (2025) ran seven experiments. In the main ones, people learned how to plant a vegetable garden or how to live a healthier life, either from Google search links or from a summary such as ChatGPT or Google's AI Overview, and then wrote advice for a friend.

The strictest test was the second experiment. Its 1,979 participants all got the same set of facts. One group got a 291-word summary with seven tips. The other group got six linked articles that covered the same seven tips, each written in the style of a different outlet. Only the format differed.

With the same facts, the summary group spent less time reading and writing

  • Web links
  • AI summary
0s50s100s150s200s250sWeb links · Reading the results · 124.3s124.3sAI summary · Reading the results · 83.7s83.7sReading the resultsWeb links · Writing the advice · 208.8s208.8sAI summary · Writing the advice · 185.2s185.2sWriting the adviceAbout a third less
Mean seconds spent reading the search results, and then writing the advice, in experiment 2. Both groups received the same facts. Rounded to one decimal place.Source: Melumad & Yun (2025), experiment 2 results
See the numbers
Web linksAI summary
Reading the results124.3s83.7s
Writing the advice208.8s185.2s

The summary group spent about a third less time on the results (32.7%, our calculation). They then reported learning fewer new things (3.71 against 3.96 on a five-point scale) and felt less ownership over what they had learned, even though the information was the same.

Why would the same facts stick less?

Because the summary did the selecting and connecting for them. The researchers argue that this effort is what deep learning is made of.

Learning from links means choosing which page to open, reading it, pulling out what overlaps, and putting the pieces together in your head. An AI summary hands you that work already done. It is easier, and you produce less of it yourself.

That fits a much older finding about memory. Information you generate yourself is remembered better than information you only read; this is called the generation effect. Bertsch and colleagues (2007) pooled 445 effect sizes from 86 studies and found an average benefit of 0.40, close to half a standard deviation.

In all three experiments, people who learned from a summary wrote shorter advice

  • Web links
  • AI summary
0306090120Web links · Exp. 1 ChatGPT · 94.694.6AI summary · Exp. 1 ChatGPT · 84.684.6Exp. 1ChatGPTWeb links · Exp. 2 same facts · 74.274.2AI summary · Exp. 2 same facts · 64.564.5Exp. 2same factsWeb links · Exp. 3 AI Overview · 90.290.2AI summary · Exp. 3 AI Overview · 65.265.2Exp. 3AI OverviewShorter with the same facts
Mean word count of the advice written for a friend after learning. Experiment 1 compared ChatGPT with Google, experiment 2 a summary and links carrying the same facts, experiment 3 Google's AI Overview with standard results. Rounded to one decimal place.Source: Melumad & Yun (2025), experiments 1 to 3 results
See the numbers
Web linksAI summary
Exp. 1 ChatGPT94.684.6
Exp. 2 same facts74.264.5
Exp. 3 AI Overview90.265.2

The advice was not just shorter. It mentioned fewer facts, and the summary group's pieces of advice were more alike. When other people read the advice without knowing how it was written, they rated the summary group's advice as less helpful and less trustworthy.

A study with students points the same way. Stadler and colleagues (2024) asked 91 university students to research nanoparticles in sunscreen with either ChatGPT or Google and then write a recommendation with reasons. The ChatGPT group had a lower cognitive load, but their reasoning and argumentation were of lower quality. In this study, though, the recommendations were not more alike.

What if the AI summary is wrong?

A summary can leave out the key point, and what it gets wrong can become what you remember.

Sim and colleagues (2026) asked two AI models, ChatGPT and Gemini, to summarize a video of a car hitting a pedestrian. Most of the summaries left out the most central event. Then they showed people the video and had them read either an accurate or a misleading AI summary. Those who read the misleading one were much less likely to remember the original event correctly: 44.8% answered the key question right, against 83.6% after an accurate summary.

Even after you have seen the original, a wrong summary can pull your memory toward it.

Source links do not fully fix this. In a follow-up by Melumad and Yun (2025), the summary came with live web links, but only 26% of participants clicked them, and the summary group still reported shallower learning.

So should you write your own summaries instead?

Not necessarily. What the research backs most strongly is not a neat summary but closing the material and recalling it.

Dunlosky and colleagues (2013) rated ten easy-to-use study techniques by whether their benefits held across ages, materials and kinds of tests.

TechniqueRatingWhy (abstract)
Practice testingHighHelps across ages, abilities, tests
Spacing out studyHighHelps across ages, abilities, tests
SummarizingLowHelps only under limited conditions
Rereading, highlightingLowDoes not consistently boost results

Summarizing was rated low utility too. It helped some students on some tests, but the conditions were limited. So replacing an AI summary with one you copy out yourself is not enough.

The difference comes from pulling the material out without looking. In Roediger and Karpicke (2006), students who restudied did better on a test five minutes later, but on tests two days and one week later the students who had practiced recall remembered substantially more. The restudy group was also more confident. The feeling that you know it all while reading a summary is the same trap. A simple way to recall on a blank page is in Is note-taking a waste of time?.

How should you use AI summaries?

Use the summary as a map before you read, and build the memory yourself after you close it. Here is an order you can try today.

  1. Read the summary first. Ask for a map, for example "list the five key questions this chapter answers". The goal is to know what to look for, not to get the answers.
  2. Read the source yourself. At least check the parts the summary points to. If you notice something the summary left out, mark it.
  3. Close everything and write three lines. Without looking at the summary or the source, write what you just learned in your own words. Where you get stuck is what is not yours yet.
  4. Compare with the AI summary. Put your three lines next to the summary and look for differences. What is missing may be on your side; what is wrong may be on the AI's.
  5. Recall it again the next day. Pick three things worth keeping and explain them without looking. Repeat a few days apart and they last longer.

Handing your memory to search is covered in The Google effect on memory.

In Aloom

Aloom does not pile up whatever AI says. You keep only the answers you choose, and each one connects to what you already know in your knowledge graph. From the next day, you pick 1, 3 or 5 a day and recall each one from a question before you see the answer. It is built to move you from reading to retrieving. Each successful recall makes that answer's ink darker. See how Aloom works

FAQ

Is it bad to use AI to summarize readings?

Not bad as such, but if the summary is all you read, you learn less deeply. In Melumad and Yun (2025), people who learned from a summary felt they learned less and wrote shorter advice. That study measured self-reported learning and what people wrote, not a memory test, and the topics were everyday ones like gardening and health. The authors note the result may differ when you only need a single fact, or in specialized fields.

Is using AI to summarize readings cheating?

That depends on your course's policy, not on research. Ask your instructor. What the research does say is that a summary which replaces the reading also replaces part of the learning, so it is worth reading the source and recalling it yourself.

Can I use AI to summarize a PDF or a lecture video for studying?

Yes. It saves time when you first skim or decide what to read. But a summary can drop the key point (Sim et al. 2026), so check anything likely to be on the exam against the source, then close it and recall it in your own words.

References