Capabilities
Summarization
Also known as: text summarisation
AI that condenses a long text into a shorter version that keeps the key points; one of the most useful everyday AI tasks.
Summarization is AI that shortens long text while keeping what matters, turning a 20-page report into a few clear bullet points, or a long email thread into a quick recap. It is one of the most practical uses of AI, saving time whenever there is more to read than you have time for. Just double-check important details, since a summary can occasionally drop or distort a key point.
Summarization condenses a longer text into a shorter one that preserves the essential information. Two styles: - Extractive: pull out the most important existing sentences. - Abstractive: rewrite the gist in new words (what modern LLMs do well).
Everyday uses
Summarising documents, meetings, research, email threads and search results. It is often a building block inside bigger systems (e.g. condensing history to fit the context window).
The main risk
A summary can introduce errors or leave out something important, including hallucinated details not in the source. For anything consequential, verify against the original.
Summarization maps a source document to a shorter faithful rendering, spanning extractive (selecting salient source spans) and abstractive (generative rewriting) paradigms; modern practice is abstractive via LLMs, often zero-shot. Beyond generic summaries, variants include query-focused, multi-document, and long-document summarisation, the last constrained by the context window and addressed by chunking, hierarchical/map-reduce strategies and retrieval.
The central quality axis is faithfulness/factual consistency: abstractive systems can introduce content unsupported by the source (a hallucination failure), so evaluation increasingly emphasises faithfulness and coverage over surface-overlap metrics (ROUGE), which correlate weakly with human judgement. It is also a core internal primitive, summarising conversation history or intermediate agent state to manage context-budget. Trade-offs balance compression ratio, coverage, faithfulness and controllability (length, focus, format).