Extractive vs. abstractive summarization

Extractive summarization builds a summary by selecting the most important sentences from the source and copying them out word for word. Abstractive summarization writes new sentences that restate the source’s meaning in fewer words. Nearly every AI meeting summary is abstractive. A language model reads the transcript and writes its own account of it, so the sentences in your meeting notes were usually never spoken by anyone.

The difference matters most when you decide how far to trust a summary, and it comes down to one line: an extractive summary can only be wrong about what it chose; an abstractive summary can also be wrong about what it says.

The two methods side by side

	Extractive	Abstractive
How it’s made	Scores each sentence in the source and keeps the top ones	A language model reads the source and writes a new, shorter text
Are the words in the summary from the source?	Yes, verbatim	Usually not. They’re paraphrase
Typical failure	Leaves out something important, or keeps a sentence that makes no sense on its own	Also states something the source didn’t say: a flattened hedge, the wrong person, a decision that wasn’t made
How visible errors are	Fairly visible, because a bad pick reads as choppy or out of context	Hard to see, because a wrong sentence reads as fluently as a right one
How to check it	Find the sentence in the source	Find the moment in the source and compare the meaning
Fit for meeting speech	Poor, because spoken sentences rarely stand alone	Good, which is why meeting tools use it
Older examples	Classic “keyword” summarizers and highlight tools	Today’s LLM-written notes, recaps and live summaries

Extractive methods are the older approach. Automatic summarization research going back to the late 1950s picked sentences by word frequency and position, and later graph-based methods ranked sentences by how central they were to the rest of the text. Abstractive summarization became practical at scale with neural language models, and it’s now the default for anything that has to read naturally.

Why meeting AI is almost all abstractive

Extraction works well on edited text such as news stories and reports, where each sentence was written to make sense on its own. Meeting speech is different, and three things about it break extraction.

• Spoken sentences don’t stand alone. People talk in fragments, restarts and fillers: “so — yeah, no, I think if we, if QA’s done by Thursday then sure.” That’s a real decision, but you can’t lift it out and drop it into a summary.
• Meaning is spread across turns. A decision often looks like this. One person proposes, a second raises a problem, a third suggests a fix, and the first says “yeah, let’s do that.” The sentence that confirms the decision is “yeah, let’s do that,” and on its own it means nothing. Only a rewrite can say “Agreed to ship Friday if QA finishes Thursday.”
• Compression has to be heavy. A one-hour call runs to thousands of words. Even the best ten sentences from it would make a poor summary, because no ten spoken sentences cover a whole meeting.

So meeting tools rewrite. That’s the right engineering choice, and it’s why AI meeting notes read well. But it changes what kind of mistakes you should look for.

What abstraction costs: the words were said by no one

Since an abstractive summary is newly written, it can include things the source doesn’t support. Researchers call this a faithfulness problem, and when the summary states something with no basis at all, hallucination. In meetings, the common failures are quieter than outright invention.

• Flattened hedges. “We could probably ship Friday if QA clears it” becomes “The team will ship Friday.” Each word is reasonable, but the condition is gone.
• Tidied-up disagreement. Two people who didn’t agree become “the team discussed options,” or worse, a consensus nobody reached.
• Misattribution. The summary has to name who said what, but it works from a transcript where speakers may only be numbered. Speaker diarization errors carry straight into “Sam agreed to…”.
• Resolved-sounding loose ends. A question that was asked and never answered can come out reading like a conclusion.

None of these look like errors. Transcription mistakes are usually visible as garbled words. Summary mistakes are fluent, confident prose. That’s the core of how accurate AI meeting notes are: the transcript fails visibly and the summary fails invisibly. Abstraction is the reason for the second half.

It’s also why accuracy metrics for the two layers don’t line up. Word error rate checks a transcript word by word against a reference. Summarization benchmarks have traditionally scored word overlap with a human-written summary (the ROUGE family of metrics is the best known), and overlap rewards copying, not being right. A summary can score well and still drop the “if”. Neither number tells you whether the meeting’s meaning came through intact.

”Extracted” usually means “written”

Meeting tools often say they extract action items and decisions. In practice those are abstractive too. The model doesn’t copy out the sentence where someone committed to something. It writes a clean line such as “Priya to send the revised quote by Friday,” pieced together from a commitment, a name and a deadline that may have come from three different turns. That’s what makes action item detection useful, and it’s also why a wrong owner or a deadline nobody stated can appear in an otherwise tidy list.

The same goes for minutes. Minutes record what was done, and the record has to be right, which is why AI can draft meeting minutes but a person should complete them and the group should approve them.

Compression ratio: why the zoom level changes the kind of summary

How abstractive a summary has to be depends on how much it’s compressing. At a typical speaking pace, as a rough illustration rather than a measurement:

• The last 30 seconds in one sentence is light compression. A sentence like that is close to a paraphrase, and there’s little room for the model to choose what to leave out.
• The last 5 minutes in a paragraph is moderate compression. The model is now choosing which points matter.
• A whole hour in two paragraphs is heavy compression. Most of the work is deciding what to drop, and that’s an editorial judgment.

This is a useful way to read a multi-resolution summary. The narrow windows are the most faithful to what was just said, so use them to answer what’s happening right now? The wide windows are an editor’s view, so use them for what has this meeting been about? and check them before you quote them. It’s also why a short window has to be summarized from its own stretch of transcript rather than cut out of a whole-call summary. Heavy compression has already thrown away the detail you need.

Hybrids and grounding

The line between the two methods is blurrier in practice. Some systems first pick out key passages (extract), then rewrite only those (abstract). Some attach a timestamp or a quote to each summary point, so you can jump to the moment it came from. That kind of grounding doesn’t make the sentence more accurate, but it makes it much cheaper to check. If a tool you’re evaluating offers it, it’s worth having. If it doesn’t, the transcript is the source of truth, and the summary is a guide to where to look in it.

How to read an abstractive meeting summary

• Check the parts where paraphrase changes meaning: numbers, dates, names, negations (“can’t” vs “can”), and conditions (“if”, “unless”, “probably”).
• Treat who-said-what as a hint. Attribution is the weakest link when the underlying transcript has numbered rather than named speakers.
• Be careful with anything that sounds settled. If the summary says “decided”, check that someone actually decided.
• Check while the room is still there. A live summary can be checked during the call, while you can still say “just to confirm — that’s only if QA clears it?” A post-meeting summary gets checked hours later, if at all, after people have acted on it. That’s the live transcript vs. live summary trade in one line: the transcript gives you the exact words, and the summary gives you the gist, which you need to check.
• Remember what a summary can’t do. It records what was said, not what was meant. For subtext and intent, ask directly.

Where Canary fits

Canary is a real-time, bot-free meeting summarizer. It captures your computer’s system audio (no bot in the call, no plugin) and shows a live, multi-resolution rolling summary — from what’s being said right now to the whole call — so you can catch up the instant your name is called.

Its summaries are abstractive. Canary transcribes the call with streaming transcription, and a language model writes each summary from the transcript. Each time window (now, about the last 30 seconds; the last 2 minutes; the last 5 minutes; and the full call) is summarized from its own slice of the transcript, at three detail levels: one sentence, a short paragraph, or two paragraphs. The widget shows one window and detail level at a time, and that’s the resolution matrix. Everything in this article applies to it.

• The narrow windows are the most faithful. “Now” in one sentence is close to a paraphrase of what was just said, and that’s the moment Canary is built for: catching up when your name is called.
• The full-call view is the most editorial, and on long meetings it’s built from the most recent part of the transcript (roughly the last 10–15 minutes), so earlier discussion may not be in it. The end-of-call decisions and action items are written the same way, from the same recent stretch.
• Summary lines don’t link back to the transcript. After the call, the review view has the full timestamped transcript, which you can search and copy. Use it to check a number, a name or an “if” before you act on it.
• Speakers are numbered, not named, so treat attribution in the summary as approximate.
• It’s cloud-processed, not on-device. The transcript is sent to a language-model provider to write the summaries. The privacy policy names the providers.

If you need the exact words, such as a quote, a verbatim record or a legal transcript, you need a transcript and probably a human check. See Canary vs. Rev for that trade, and Canary vs. Tactiq for live transcript vs. live summary. If you need to know where the conversation is right now, an abstractive summary you can glance at and check before anyone leaves is the better tool. For the whole pipeline from audio to summary, see how AI meeting notetakers work.