Automatic summarisers fall into two families that work in fundamentally different ways. The distinction is not academic — it determines whether the summary can contain a claim your source document did not make.
Extractive: selecting sentences
An extractive summariser scores each sentence in the original and returns the highest-scoring ones, unchanged. Classical approaches score on term frequency: words appearing often across the document are treated as topical, and sentences dense in those words are treated as central. Position is commonly weighted too, since most writing states its point early.
The defining property is that every sentence in the output appears verbatim in the input. The summariser cannot introduce a fact, because it cannot write anything. Its failure mode is a summary that reads choppily, or one that pulls a sentence containing a pronoun whose referent is now missing.
Abstractive: generating new text
An abstractive summariser generates new sentences, the way a person would. Modern versions are transformer models trained on large collections of documents paired with human-written summaries, learning to produce condensed text conditioned on the input.
The output reads far better. It can combine information from separate paragraphs into one clean sentence, and it does not suffer dangling references. This is what people usually mean by AI summarisation.
The hallucination problem
Because an abstractive model generates text rather than selecting it, it can produce fluent statements that the source does not support. This is well documented in the research literature and is not a bug that better training has eliminated — it is inherent to generating text conditioned on a document.
The failure is dangerous precisely because it is invisible. A hallucinated summary does not look wrong; it reads smoothly and confidently, and you would only catch it by checking against the original — which defeats much of the point of summarising.
Choosing between them
For anything where accuracy matters more than elegance — legal documents, medical information, financial reporting, research you intend to cite — extractive is the safer choice, because you can trust that every sentence genuinely appeared in the source.
For getting the gist of a long article you would otherwise skim, abstractive reads better and the occasional imprecision costs little.
What summarisers do badly regardless
- Dialogue and interviews, where meaning is distributed across an exchange rather than concentrated in sentences.
- Narrative, where sequence carries the meaning and removing sentences destroys the thread.
- Documents whose point is a single buried caveat rather than a repeated theme — frequency-based scoring will discard exactly that sentence.
- Bulleted or fragmentary text, which has no sentence structure to score.