What summarisation is good at
Modern models are genuinely strong at extracting structure and main claims from long text. For orientation — what is this chapter about, how is it organised, which sections matter for my question — a summary is fast, accurate enough, and saves real time.
- Deciding what to read. A summary of a 60-page chapter tells you which 15 pages you actually need. This is the highest-value use and it's underrated.
- Recovering a chapter you read weeks ago. Reactivating existing knowledge is a different task from acquiring it, and summaries do it well.
- Getting the structure before a dense read. The survey step, automated — see SQ3R.
- Comparing two sources' treatment of the same topic, where you want the shape of the disagreement rather than the detail.
- Catching up on a lecture you missed, well enough to follow the next one while you arrange to cover it properly.
What it drops, and why that's the exam
| Compressed away | Why it matters |
|---|---|
| Conditions and caveats | "This holds only for small samples" is the qualification that separates band descriptors. |
| Worked examples | The example is where the method becomes reproducible. A stated method isn't a usable one. |
| Counterexamples and exceptions | Examiners target exceptions precisely because summaries drop them. |
| The reasoning between claims | Summaries keep conclusions and drop the argument. Exams mark the argument. |
| Emphasis and proportion | A summary flattens: three lines on the central mechanism, three on a footnote. |
| Difficulty | A hard idea made to sound simple feels understood. That feeling is the problem. |
The last row is the subtle one. A good summary makes difficult material feel easy, and that fluency is indistinguishable from comprehension while you're reading it — the same illusion documented in rereading versus retrieval, with the compression making it stronger rather than weaker.
Accuracy: what actually goes wrong
Outright invention is rarer in summarisation than in open-ended generation, because the source text is right there. The realistic failure modes are quieter.
- Dropped qualifiers. "X causes Y" where the source said "X is associated with Y in observational studies". This is the single commonest and most costly error.
- Flattened uncertainty. A contested claim presented as settled, because summaries prefer declarative sentences.
- Merged distinctions. Two similar concepts the chapter carefully separated, combined into one paragraph.
- Confident coverage of a skipped section, where the summary implies completeness it doesn't have.
- Blending in general knowledge that isn't in your chapter at all — invisible unless the summary cites where each point came from.
Using summaries without paying the price
- 1
Summarise to triage, then read the parts that matter
Use it to find the 20% of the chapter that's examinable, then read that 20% properly. This is the workflow that's strictly better than both alternatives — reading everything, or reading only the summary.
- 2
Demand citations, and spot-check three claims
If each point links to a page, check three at random against the source. Thirty seconds, and it catches dropped qualifiers, which is where the damage is.
- 3
Ask specifically for what summaries drop
"List every condition, exception and caveat in this chapter" is a different and more useful request than "summarise this chapter". Ask for both.
- 4
Write your own summary afterwards, from memory
This converts a reading exercise into retrieval. The AI summary is then a checking tool rather than the thing you studied, which is the correct role for it.
- 5
Turn the summary into questions, not into notes
A summary you re-read decays like any other text. A summary converted into questions and reviewed is retained — see spaced repetition.
Where a summary is the wrong tool entirely
Some material resists compression by its nature, and asking for a summary produces something that reads well and teaches nothing.
- Mathematical derivations. The steps are the content; a summary of a proof is a statement of the theorem.
- Legal cases, where the reasoning is the holding — see how to read a legal case.
- Primary sources you'll quote. You need the language, not the gist.
- Anything where you'll be asked to apply a method. Applying requires the worked example that a summary removes.
- Material you've never encountered at all. Summaries are excellent for compressing the known and poor for introducing the unknown.
The honest bottom line
If you have twelve chapters and time for four, summarise all twelve to decide which four, then read those four properly and convert them into questions. That beats both the exhaustive read you won't finish and the summary-only approach that leaves you fluent about material you can't actually use.
What summaries can never do is the retrieval. However good the compression, reading it is still reading — and the thing that decides what you'll know in the exam is what you produced from memory, not what you consumed. How to remember what you read is the same argument applied to the unsummarised version.