Key points
- Summarizing isn't shrinking a piece of writing — it's choosing what to keep. Judgment enters the moment you decide what to leave in.
- There are two broad approaches: pulling original sentences out as-is, and reading it and writing new sentences.
- The same piece of writing gets summarized differently depending on who's reading it and why.
- A document too long to fit in one pass gets summarized in chunks, then those summaries get summarized again.
- Numbers, deadlines, conditions, and the word "not" are what go missing especially often.
The analogy
After a two-hour meeting, someone writes a one-page notice for whoever couldn't make it. Copy down everything that was said and that's not a notice anymore — that's a transcript. Whoever writes the notice isn't recording everything said; they're deciding what to cut.
Change who's reading the same meeting and what survives changes too. A notice sent to teammates keeps everyone's assignment and deadline; a report sent up to an executive keeps only the conclusion and the budget. The side that got cut isn't wrong — it's just written for a different reader.
A good set of minutes isn't good because it's short — it's good because nothing that got cut ever needs recovering. Summarizing works the same way. Length isn't what decides quality; what's missing is.
In detail
It's a choice, not a compression
Ask for a summary and it's easy to picture an AI compressing the text. What actually happens is a judgment call about what to keep and what to drop. Keep two paragraphs out of ten, and the other eight just got decided to matter less.
That judgment comes from what it learned during training — pairs of countless human-written pieces alongside their own headlines, conclusions, and abstracts, building a sense of "in writing like this, this is usually the part that comes first."
That's why an AI's summary tends to be reasonably sound but not necessarily what fits you. If what you actually wanted today wasn't the conclusion but the counterargument, a generic summary cuts exactly that part.
Extracting versus rewriting
Summarizing splits into two long-standing approaches. One pulls sentences straight out of the original that look important and strings them together. Nothing gets added that wasn't in the original, but the sentences can feel disconnected, and a pronoun's referent sometimes disappears.
The other reads the whole thing and writes new sentences from scratch. It can pull a story scattered across several paragraphs into one sentence, which reads far better — but wording that wasn't in the original can slip in during the rewrite.
Most summaries an AI service produces today lean toward rewriting. For writing where exact wording matters — a contract, a set of medical instructions, anything you'll need to check against the original — it's safer to ask for extraction, with the source sentences shown alongside. Even so, a sentence the AI shows you may itself be reconstructed, so confirming it really appears in the original means looking it up in the original.
You need to say who's reading
"Summarize this" by itself is like asking for meeting minutes without saying who they're going to. Tell it who's reading and what it's for, and the result changes considerably.
Length is worth setting too. Three lines or one paragraph, a list or flowing prose — how much information survives shifts with the choice. Specifying a shape, something like "split what's decided from what's still open," cuts down on things going missing.
It's fine if the first attempt isn't great. Point out "the budget part is missing" and the next summary comes back noticeably better.
A long document gets folded twice
There's a hard limit on how much text an AI can take in at once. Something very long, beyond what fits in one pass, won't go in whole, so it gets split into a front, middle, and back section, each summarized separately, and those summaries get summarized again. Even when it fits within the limit, a very long text makes it easy to miss the middle, so it's sometimes split on purpose.
This approach makes long documents workable, but it costs something twice over. Whatever got dropped in the first pass has no chance of making it into the second. Anything that spans a long stretch — a condition set in chapter three that gets reversed in chapter nine — is especially likely to vanish.
Where a chunk gets cut off matters too. Cut in the middle of a paragraph and that paragraph's conclusion never survives intact in either piece.
Certain things go missing on cue
The things that disappear especially often from a summary are numbers, deadlines, conditions, and the word "not." "The budget was approved, but spending doesn't start until Q3" shrinking down to just "budget approved" is a textbook case.
Drop a negation and the meaning can flip outright. "Not recommended" turning into "recommended" sends a reader moving in exactly the wrong direction. When a conclusion carries a condition, it's worth asking to see that original sentence alongside the summary.
More precisely
An AI's summary is closer to a new piece of writing continued from the original than a genuine compression of something it understood. Since it's built the same way as picking the next word, sentences absent from the original can slip in naturally. Making up content that wasn't in the source and presenting it convincingly is called hallucination.
The analogy has a limit. Whoever writes meeting minutes sat in the room and picked up on context and mood; an AI only sees the text it was handed. Anything agreed on with a glance rather than said out loud, anything outside the document, can't make it into the summary. A person also remembers what they cut, while an AI keeps no separate list of what it dropped — finding out what's missing means going back to the original either way. The long-standing automated scores used to grade summary quality check how much a summary overlaps with a human-written one, so they're weak at telling a smooth summary apart from an accurate one. Newer scores that compare directly against the source exist, but they aren't perfect yet.
Try it yourself
- Lilys AI (video and document summaries) ailearn.space Feed it one long video and check the summary against the original to see what stayed and what got cut
- NotebookLM ailearn.space Ask for "just the conclusion" and then "just the action items" from the same material and watch the summary shift with the goal
Common misconceptions
It's easy to think summarizing is a mechanical job of just cutting length, but actually it's a judgment call about what to drop, so the result should change when the goal changes.
It's easy to assume a summary only contains what was in the original, but actually the rewriting process can let in wording or figures that weren't there.
It's easy to think a smooth summary is a good one, but actually a summary reads smoothly even with an entire condition or deadline missing, so smoothness alone can't tell you what's gone.
One-line summary
Summarizing isn't compressing a long piece of writing shorter — it's deciding who's going to read it and choosing what to drop.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17