Why do I let AI summarize what I should read myself?

Because a summary answers the question you never actually asked. You wanted to know whether you understood the thing, and you got a fluent paraphrase instead. Reading a clean summary produces the same ease you feel when you do understand something, so the shortcut collects a yes it did not earn.

Psychologists call the judgement you make at the end of a document a judgement of learning, and its honest form is a retrieval test: close the file and say what it argued. Koriat's cue-utilization work in 1997 found readers doing something cheaper. They estimate what they know from how easily the material read rather than from what they can retrieve.

A summary manufactures exactly that feeling: short sentences, clean structure, no loose ends. It also removes the thing that produces retrieval, because the summary already did the pulling. A confusing original sets off an alarm that makes you slow down and reread, and a smooth summary never does, so the sign that understanding is missing shows up days later.

The 2026 calibration evidence points the same way. A metacognition study in the Journal of Intelligence measured comprehension and confidence in fluent readers and found fluency going with larger calibration error: more fluent readers tended to overestimate their comprehension. Those readers were fifth graders, and their cue is the one everybody uses.

Why does a summary change what you remember of the source?

Because a summary does not leave a gap where the source was. It takes the source's place. In a 2026 experiment, readers of a misleading summary were right about the manipulated detail 44.8 percent of the time, against 83.6 percent of readers given an accurate summary of the same event.

That is the hardest thing the research currently says, and it is a different claim from "you missed a detail". In Sim and colleagues' 2026 AIES study, 331 people watched a video of a car hitting a pedestrian and then read an AI-written summary of it. The summary either matched the video or quietly swapped one detail, whether the car had stopped at a stop sign or a yield sign. Then came a memory test on the original event.

The summary did not merely leave the source unread. It became the memory of it, and the effect held whether or not participants were told the summary was AI-written and whether or not they trusted AI. The same paper's audit found the omissions that matter most: most summaries of that video left out its central event. One video study is not a law of nature, and reading a report is not watching an accident, but this is the strongest evidence on summaries and memory that exists.

Is reading a summary the same as reading the thing?

No, and the 2026 evidence sharpens the difference. You do not just miss details. Summaries become the material. Fu and colleagues' eight-week study of 15 undergraduates found students treating AI summaries as the primary text and only then deciding which sections deserved the original, a pattern the authors name reading through AI rather than with it.

That eight-week study tracked 838 prompts across 239 reading sessions. Most sessions contained exactly the three prompts the assignment required and stopped there, and students named efficiency as the driver. Nothing in the logs looks like triage. It looks like replacement.

You are not using the summary to decide what to read. You are using it to decide you have read.

The correlational work says the cost is real without settling its size. A 2026 Frontiers in Psychology survey of 608 undergraduates found heavier use of summarization tools going with lower deep reading comprehension, working through less elaborative processing, the connecting and questioning you do with a text. The negative indirect effect was present at low and average reading motivation and disappeared at high motivation. The design is cross-sectional, so it cannot say whether summaries caused the gap or weaker reading led to them. It is not the tool that decides. It is what the tool replaces.

How do I keep the reading when I use a summary?

Generate before you open. Write three lines from memory on what the document argued, or list the keywords you would search for, and only then read the summary. The comparison is the point, because a mismatch between your version and the model's is a free comprehension test.

The keyword-generation work by Thiede and colleagues found that listing a few keywords before you judge your own understanding improves that judgement's accuracy, which is the part a summary quietly inflates. You are not trying to beat the model. You are trying to keep a measurement of yourself the model cannot make.

Keep the two lanes summaries are good for: deciding whether something is worth reading, and refreshing recall after you have read it. Never for the text you will be asked to defend. And do not take the padding verdict at face value. In a thread about reports nobody reads, a business analyst described the arc exactly: weeks of drafting, then the five bullets made from it, and the bullets are all anyone sees. When the summary is the only text in circulation, nobody left in the loop can notice what it dropped.

Then see what Absorb scores in a short daily round of reading, listening, and analyzing, ten to twelve minutes at a time.