Extractive vs abstractive summaries: how each works and when to trust it

Summaries now come from two very different machines. The older kind picks sentences out of a text; the newer kind writes new ones. They fail in opposite ways, and knowing which you're reading decides how far to trust it. The text summarizer on this site is the first kind, by design; here's the reasoning.

Two kinds of summary

An extractive summary is a subset of the original's sentences, in their original order — a highlighter, automated. An abstractive summary is new text that paraphrases and compresses — what a person writes, and what large language models produce. Extractive summaries are guaranteed faithful and often clunky; abstractive summaries are fluent and sometimes wrong. Research systems since the 1950s (Luhn's 1958 work at IBM is the usual starting point) were extractive because writing was beyond them; the last few years made abstractive summaries cheap, which is why the distinction suddenly matters to everyone.

How extractive summarisation works

Score every sentence, keep the best, present them in order. The scoring is where methods differ. Frequency-based scoring (Luhn's idea) counts the document's content words — stopwords like "the" removed — and rates a sentence by how many frequent terms it contains, normalised for length so long sentences don't win by default. Position matters: first sentences of a document and of paragraphs carry more, because writers front-load. Graph methods like TextRank treat sentences as nodes linked by similarity and rank them the way PageRank ranks web pages — a sentence that resembles many others is central. All of these run in milliseconds on a device, need no training data, and cannot say anything the source didn't. Their weakness is coherence: the sentences don't join up, pronouns lose their referents, and the important point may be spread across three sentences none of which scores well alone.

How abstractive (AI) summarisation works

A language model reads the text and generates a summary token by token, conditioned on the input and on everything it learned in training. The result reads like a person wrote it and can compress ideas rather than sentences — "the report recommends three changes" where the source lists them across pages. The failure mode is confident invention: a statistic that isn't in the source, a recommendation the author didn't make, a name swapped — hallucination — and it is undetectable without reading the original, which defeats the purpose. Quality varies with the model, the length of the input and the domain; the best models hallucinate rarely on short, clear inputs and more on long or technical ones.

The trust question

Ask what the summary is for. For deciding whether to read something, either kind works. For quoting, citing, briefing a decision or summarising a contract, extractive is safer: every sentence is verifiably in the source. For readability — a newsletter blurb, a quick gist of a long thread — abstractive is nicer and the stakes are low. The failure to avoid is the hybrid people commonly practise: reading an AI summary of a document they never open and acting on it as if they had. A summary is a map; on anything that matters, the territory is the document. The words "extractive" and "abstractive" are worth knowing precisely so you can ask which a tool is.

Using summaries well

• Set the length by purpose: 10–15% for a gist, 25–30% for a working summary, more when every point matters.
• Prefer bullet output for extractive summaries; the disjointedness reads as a list rather than a broken paragraph.
• Check the key terms the summariser reports against what you expected the document to be about — a mismatch means it found a different topic than you did.
• For narrative and dialogue, summarise by hand; keyword-dense sentences aren't the important ones there.

The readability checker grades the summary, the word frequency counter shows the terms it scored, and the keyword density checker the phrases.