Abstractive summarization generates a summary by rewriting source material in fresh language — synthesising ideas, compressing detail, and producing output that reads like a human wrote it, not a highlighter extracted it.
Why it matters
A language model asked to summarise a document doesn't copy sentences; it reads the content and produces new text that captures the gist. The result is fluent and often easier to read than a list of lifted quotes. That's the appeal: abstractive summaries work the way human note-taking does — understanding first, then expression.
The risk is the flip side of that same capability. Because the model is generating new text, it can drift from the source. Small paraphrases can shift meaning; low-confidence passages can be filled in with plausible-but-wrong detail. That's hallucination in its most dangerous form, because the output looks completely natural.
The fix is source-grounding: constrain the model to answer only from retrieved passages, and attach a citation to each claim. Grounded abstractive summaries give you the readability of rewritten prose with the auditability of extracted text — each fluent sentence still has a source sentence it can be checked against.
How it compares
Extractive summarization stays closer to the original wording, which makes it inherently safe from hallucination but often choppy to read. In practice, the most useful document AI blends both approaches: retrieval selects the right passages, abstraction rewrites them cleanly, and citations keep every claim traceable. Sidenote's summaries follow exactly this pattern.