Why picking sentences is not summarising
Most free summarisers are extractive. They count how often each word appears, score every sentence by how many of those frequent words it holds, rank the sentences and hand back the top ten percent in their original order. Nothing is written. The output is the source with holes cut in it.
You notice it at the joins. A pulled sentence opens "This meant the committee rejected the proposal" when the committee appeared two paragraphs earlier and the proposal three pages before that. Pronouns point at nothing, figures arrive without the units they were defined in, and you end up reconstructing the context that was deleted.
This one is abstractive. It works from the whole passage and writes fresh sentences, so four paragraphs building towards a single finding come back as one clause stating that finding, and a table of results becomes the two rows that moved. The wording is new — which is the point, and also the risk.
Why length changes what the summary is for
Length is not a volume dial on the same output. At ten percent a forty-page report becomes four paragraphs, and that forces a decision about what the report was actually for — the finding, the recommendation, the one number that changed. Everything holding it up goes. You get a verdict, not an account.
At forty percent almost nothing has to be dropped, only tightened. The method survives, the caveats survive, the three exceptions the author flagged survive, and what comes back reads like the same report written by somebody in a hurry. That is the version you want before a viva, or before a client meeting where you will be asked follow-up questions.
So choose by what you are about to do with it. Ten percent to decide whether the document deserves an afternoon. Twenty-five to brief a colleague who will never open it. Forty when you have to defend the contents. Shape is a separate choice — bullets when you are hunting for the items that need action, prose when the reasoning has to hold together.
Where a summary stops being a substitute
Here is the constraint worth stating plainly. A summary of something you have already read is checkable, because you notice at once when a caveat has gone missing. A summary of something you have not read cannot be checked by you — and that is precisely the case you wanted the tool for.
The specific failure is compression of a hedge. "Enrolment rose in three of eleven districts" can flatten to "enrolment rose", and those are different claims. Names, dates and rupee figures are carried through deliberately for this reason, but the qualifier attached to a figure is easier to lose than the figure itself.
Treat the output as a reading aid rather than a source. Where accuracy carries weight — a citation in your dissertation, a clause in a lease, a figure going into a board pack — open the original at the relevant passage and read that passage. Finding the passage is the job this does well.