How AI Text Summarisers Work (and When to Trust Them)

An AI summariser reads a document and writes a new, shorter passage that captures its main points, using a language model trained to condense text rather than simply copying the first or most-repeated sentences. A good one on a report or article keeps the key facts and conclusions while leaving out examples, background and repetition, and lets you choose how short the result should be and whether it reads as a paragraph or a list of points.

Key takeaways

Abstractive summarisers write new sentences; extractive ones just select existing sentences. Each has trade-offs.
A summary this short cannot include everything: it keeps what the model judges most important, which will not always match what you need.
Long documents are usually summarised in sections and then combined, which can lose connections between sections far apart in the text.
A summariser that runs in your browser never uploads the document; one that calls a cloud API does, even if the provider promises not to store it.
Always check a summary against the original before relying on a specific number, date, name or quote.

Extractive vs abstractive summarisation

There are two different approaches, and knowing which one a tool uses explains a lot about its output.

Extractive summarisation picks a handful of existing sentences from the document and presents them as the summary, usually the ones with the most important-looking or most central words. It can never misquote the text, because every sentence is copied verbatim, but it reads as a disconnected list and misses anything that needed rewording to make sense out of context.
Abstractive summarisation uses a language model to write new sentences that paraphrase and combine ideas from the document, the way a person would summarise it in their own words. It reads naturally and can combine information from different parts of the text, but because it generates new wording, it can occasionally soften a nuance, drop a secondary detail, or very rarely phrase something in a way that drifts slightly from the source.

Our PDF & Word Summariser is abstractive: it uses a transformer model (DistilBART, trained specifically for summarising news and reports) running in your browser, not a cloud AI service.

How a long document gets summarised

Language models can only read a limited amount of text at once, typically under a thousand words for a compact model. A long report will not fit in a single pass, so a practical summariser:

splits the document into sections at sentence boundaries, so no sentence is cut in half;
summarises each section on its own;
combines the section summaries and summarises that combination once more, so the final result reads as one piece rather than a row of separate mini-summaries.

This map-then-combine method scales to books and long reports, but it has a real limit: a connection between a fact on page 2 and a fact on page 40 can be lost if they end up in different sections and neither section summary mentions the link. For documents where every cross-reference matters, such as a contract, read the full text; use the summary to decide where to look closer, not as a replacement for reading it.

Why a summary can be wrong, even without "hallucinating"

People sometimes assume a summariser is accurate because, unlike a chatbot answering open questions, it only has to describe text that is right in front of it. In practice, three things still go wrong:

Omission presented as absence. Leaving out a detail is the summariser's job, but a reader can mistake "not mentioned in the summary" for "not in the document."
Compression errors. Merging two related but distinct sentences can create a new sentence that overstates a conclusion, drops a condition ("usually" becoming unqualified), or attaches a number to the wrong item.
Weak input. Tables, code, reference lists, bullet fragments and very short text give a language model little genuine prose to work with, so the output can repeat itself or state the obvious.

None of this means AI summarisers are unreliable for their actual purpose, which is a fast first read of a document you have not read yet, or a refresher on one you have. It means the summary is a starting point, and anything you plan to act on, cite or quote should be checked against the source.

Where does the document go?

Many free summarisers send your document to a server, because running a language model used to require more computing power than a browser could offer. That is no longer always true: small, efficient models can now run inside the browser itself using WebAssembly, so the text never has to leave your device. The trade-off is a one-time download of the model (ours is about 275 MB) and slower processing than a large cloud model on a server, in exchange for documents, including contracts, medical letters and anything else sensitive, never being uploaded anywhere.

Getting a better summary

Match the length to the use. A short summary for a quick decision on whether to read something at all; medium for most documents; long when you specifically want more of the detail kept.
Clean the input first. If a PDF imports with broken line breaks, run it through PDF to Text and check the result reads sensibly before summarising it.
Summarise prose, not forms. Tables, order forms and reference lists do not summarise well; the useful content there is the specific values, not a condensed description of them.
Re-read anything you will rely on. Use the summary to find the relevant section, then read that section in full before quoting a figure or date from it.

Frequently asked questions

Are AI summaries accurate?

They are a reasonable approximation of a document's main points, not a guaranteed-accurate record of it. Compression can soften a nuance or misattribute a detail, and a summary can never include everything, so check the original before relying on a specific fact.

What is the difference between extractive and abstractive summarisation?

Extractive summarisation selects existing sentences from the text; abstractive summarisation writes new sentences that paraphrase it. Extractive is safer against misquoting but reads less naturally; abstractive reads better but can occasionally reword something imprecisely.

Can an AI summariser handle a whole book?

Yes, in principle, by summarising it section by section and then combining those summaries, though connections between widely separated sections can be lost, and processing a very long document takes correspondingly longer.

Is it safe to summarise a confidential document?

Only with a summariser that processes the document on your own device. Many free tools upload the document to a server first; check that before using one for anything sensitive.

Why did the summary repeat itself or state the obvious?

This usually means the input had little real content to condense: a short note, a table, a reference list or disconnected bullet points. Summarisers work best on connected prose of a few hundred words or more.

Tutti gli strumenti