Black October, explainedAppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

Summarisation

Also called summarization, AI summary, summarize

Summarisation is AI shortening a long text, like an article, report or meeting, into a few lines that keep the main points.

Inside language modelsEveryday

In plain words

We summarise all the time. A friend asks how the film was, and we give the plot in three sentences, not three hours.

AI summarisation does the same for text. There are two basic ways. The older way picks out the most important sentences and copies them as they are, like using a highlighter. The newer way reads the whole thing and writes fresh, shorter sentences in its own words, like a student writing notes.

Today's chatbots and email, meeting and news apps mostly use the second way, powered by large language models. It reads more naturally, but it brings a risk: the summary can state something the original never said. For anything important, check the summary against the source.

Everyday example

The family WhatsApp group after a day away

We come back to 300 unread messages in the family group. A cousin sums it up: "Wedding moved to the 14th, Nani is fine after the check-up, and someone needs to book the hall." Three lines replace 300 messages. A good summary keeps the decisions and the to-dos, and drops the forwards and good-morning stickers.

Where the example breaks: Our cousin was in the chat and knows what matters to us. The AI only has the text, so it may give weight to the wrong thing, miss a joke, or mix up who said what.

Go deeper

How it actually works

Summarisation is one of the oldest tasks in natural language processing. Researchers split it into extractive and abstractive methods.

  1. Extractive: pick sentences

    In 1958, IBM researcher Hans Peter Luhn had a computer count which meaningful words appeared most often in an article, score each sentence by how many of them it held close together, and print the top sentences as an "auto-abstract". Extractive methods are still used because they never invent wording.

  2. Abstractive: write new sentences

    Neural encoder-decoder models from the mid-2010s learned to write summaries in new words. They tended to repeat themselves and garble facts. A 2017 model by Abigail See and colleagues added the option to copy words straight from the source, which helped with names and numbers.

  3. Language models summarise on request

    Today a large language model can summarise from a plain instruction (a prompt) with no special training. It can also follow requests like "in 5 bullet points" or "for a 10 year old". Long documents must fit in its context window, or be split into parts and summarised in stages.

  4. The faithfulness problem

    A 2020 Google study found that people judging summaries from neural models spotted made-up content in summaries from every model they tested. These invented details are a kind of hallucination. In January 2025 Apple paused its iPhone notification summaries for news apps after the BBC complained that summaries of its alerts got facts wrong.

  5. Measuring a summary

    A common automatic score, ROUGE (2004), counts how many words and word sequences overlap with summaries written by people. It is fast but cannot tell whether a summary is true, so careful teams also check summaries by hand.

Early automatic summary
Hans Peter Luhn, IBM, "The Automatic Creation of Literature Abstracts", 1958
Two kinds
Extractive (copy key sentences) and abstractive (write new ones)
Main risk
Saying things the source does not say; check names, numbers and dates
Common score
ROUGE, Chin-Yew Lin, 2004: overlap with human-written summaries

Questions people ask

Can I trust AI summaries?

As a quick guide, often yes. As a final record, no. AI summaries are usually right about the main thrust but can drop a key caveat, mix up numbers or add a detail that is not in the source. For contracts, medical or money matters, read the original.

What is the difference between extractive and abstractive summarisation?

Extractive summarisation copies the most important sentences word for word, like a highlighter. Abstractive summarisation writes new sentences in its own words, like taking notes. Abstractive reads better but is more likely to introduce errors.

How do AI tools summarise long PDFs or books?

If the text fits in the model's context window, it reads it all at once. If not, the tool splits it into chunks, summarises each, then summarises the summaries. Some tools instead search for the relevant parts first, a method called retrieval-augmented generation.

Why do AI summaries sometimes miss the most important point?

The model decides importance from patterns, not from knowing what you care about. Research has also found models use information at the start and end of long inputs better than in the middle. Telling the AI what you need the summary for helps a lot.

Is it summarisation or summarization?

Both are correct. Summarisation is the British and Indian spelling; summarization is the American one. They mean exactly the same thing.

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked October 2026.