What is the difference between extractive and abstractive summarization?

Extractive summarization picks and copies sentences straight from the source: it cannot invent anything, but the result can read choppily. Abstractive summarization rewrites in its own words: it reads better, but can introduce claims the source never made.

Can an AI summary make things up?

Yes — that is the main risk with abstractive methods, and the term for it is hallucination. The practical defence is to require every claim in the summary to be tied to a specific passage in the source, verifiable in one click.

How is summary quality measured?

The classic ROUGE metrics, introduced by Chin-Yew Lin in 2004, measure n-gram overlap with a reference summary. They are now considered insufficient on their own and are paired with factual-faithfulness checks and human judgement.

What is automatic summarization?

Automatic summarization is the generation, by software, of a condensed version of a text or transcript that preserves its essential information. Extractive summarization selects sentences from the document; abstractive summarization rewrites the content in new words.

What exactly are we talking about?

Automatic summarization is the machine generation of a short version of a document that preserves whatever information is judged essential. It is one of the founding tasks of natural language processing: Hans Peter Luhn laid the groundwork at IBM in 1958 with a method that scores sentences by the frequency of the significant words they contain.

The success criterion is not brevity but preservation. A ten-line summary that drops the document’s conclusion is worse than a thirty-line one that keeps it. Which is why compression ratio, on its own, measures nothing.

Extractive or abstractive: what changes?

Criterion	Extractive	Abstractive
Principle	Selects existing sentences	Rewrites the content
Risk of invention	None	Real
Readability	Sometimes choppy	High
Traceability	Immediate	Must be reconstructed
Typical methods	TextRank, LexRank	Language models
Best used for	Legal and technical documents	Transcripts, long-form content

TextRank, published by Rada Mihalcea and Paul Tarau in 2004, remains the extractive reference: it applies PageRank-style logic to a graph of the document’s sentences. Abstractive approaches shifted with the Transformer architecture introduced by Vaswani and co-authors in 2017, then with large language models from 2020 onward.

How do you measure a summary’s quality?

Three dimensions, routinely conflated, need to be assessed separately.

Coverage: does the summary contain the document’s salient points? This is roughly what the ROUGE metrics, proposed by Chin-Yew Lin in 2004, approximate by comparing the produced summary’s n-grams against a human reference.
Factual faithfulness: is every claim in the summary supported by the document? A perfectly fluent and perfectly false sentence still scores well on ROUGE if it reuses the right vocabulary.
Usefulness: does the summary let you decide whether to read the source, or to skip it? This is the criterion that matters in real use, and the only one no automatic metric captures.

What are the typical failures?

Hallucination. An abstractive model can produce a date, a figure or an attribution that never appears in the source. It is getting rarer but has not gone away, and it is all the more dangerous because the output reads convincingly.

Losing the disagreement. Summaries tend to smooth over contradictions. An article laying out two opposing positions often comes back as one averaged position neither side would defend.

Flattening the hierarchy. With no signal about structure, a model can give an offhand remark as much room as the main conclusion.

Dependence on input format. An unpunctuated podcast transcript with unidentified speakers degrades the result badly. Summary quality is capped by transcript quality.

Losing the source. A summary detached from the document it came from cannot be interrogated. The moment a question gets more specific than the summary, there is nowhere to go.

How do you make a summary trustworthy in practice?

Control the input. A transcript with punctuation and identified speakers produces a markedly better summary than the same audio processed raw.
Demand traceability. Every claim should point back to its source passage. It is the only verification that holds at scale, because it takes seconds.
Match length to the document. A fixed ratio is a mistake: a one-hour interview and a dense ten-page report do not carry the same information density.
Keep the source. A summary with no route back to the original document is a dead end the moment the question gets specific.
Check numbers and proper nouns. They carry the most information and are the most frequently mangled.

What makes summaries useful day to day?

A single summary is worth little: it saves one reading, once. Value shows up when summaries accumulate in one space and become comparable — when you can see that a notion recurs across six sources and that two of them contradict each other.

That is the logic Synthiz follows: sources — articles, videos, podcasts, PDFs, RSS feeds — are transcribed then synthesised with sourced quotations tied to the exact passage, and poured into a thematic "memory" where shared concepts link from one source to the next. The summary stops being a disposable deliverable and becomes a verifiable piece of a larger whole.