AI text summarization uses two genuinely fundamentally different approaches — extractive and abstractive — and understanding the real difference explains why some summaries read as disjointed while others read as genuinely fluent prose.
Extractive Summarization: Selecting Existing Sentences
Extractive: identifies and selects the most IMPORTANT EXISTING
sentences from the source text, verbatim, and combines
them into a summary - genuinely never generates new wording
Extractive approaches score each sentence in the source text by importance (often using techniques like TextRank, a graph-based ranking algorithm related to Google’s original PageRank), then select and assemble the highest-scoring sentences — genuinely guaranteed to be factually accurate to the source (since it’s using the source’s own actual words), but can read as somewhat disjointed since the selected sentences weren’t originally written to flow together as a standalone summary.
Abstractive Summarization: Generating New Text
Abstractive approaches (what most modern LLM-based summarization actually does) generate genuinely new sentences that capture the source’s meaning, potentially using entirely different wording than the original — reads as far more natural, fluent prose, but carries real risk of subtly misrepresenting or “hallucinating” details not actually present in the original source text.
The Practical Trade-off
Extractive summarization is the safer, lower-risk choice for genuinely high-stakes accuracy requirements (legal or medical document summarization, where a fabricated detail could be genuinely consequential) — abstractive summarization reads far better for general use, but its outputs genuinely benefit from spot-checking against the original source, especially for specific numbers, names, and dates that are easy for the model to subtly alter without anyone immediately noticing.