What is abstractive summarization?
Abstractive summarization produces a summary in new words: it paraphrases and condenses the source rather than quoting it, much as a person would sum up an article for a colleague.
Abstractive text summarization is usually done with large language models, which generate the summary one word at a time.
An abstractive summarization example
Given a long article announcing that a product’s price rises from $29 to $39 for new customers only, an abstractive summary might read: “Prices are going up for new sign-ups; existing customers keep their current rate.” It is shorter and smoother than any sentence in the original, and none of its words had to come from the source.
Abstractive summarization hallucination
Because the model writes new text, it can state things the source never said: a wrong figure, an invented quote, a conclusion that merely sounds plausible. This is called hallucination. The risk is highest with numbers, names and technical detail, which is exactly where a wrong summary does the most damage.
When to use abstractive summarization
Use it where readability matters and a person will check the result: internal notes, first drafts, catching up on long threads. Avoid it, or check every claim against the source, for anything published under your name, especially quotes and figures. For those, extractive summarization is the safer default.
Last reviewed 2026-11-01.