Post #216851

2026-02-01 14:11 UTC

I ought to write this up properly, but here's a tl;dr in advance: #AI does not *summarise* documents - it *shortens* them by using structural similarity to documents in its training data to guess/predict which sentences are important and which are not. If that is what you want, fine. But don't be fooled into thinking it is doing anything like human summarisers do. They work on the *meaning* of the document in front of them.

Replies (3)

@carsten@minnesotasocial.net 2026-02-01 15:19

@tomstoneham@dair-community.social an excellent description, thank you!

Open ##2380409
@whitneymcn@mastodon.xyz 2026-02-01 15:33

@tomstoneham@dair-community.social Full disclosure: I summarized depositions for a living for a number of years, which shapes my opinion here, but I totally agree. Good summarization is like an efficient software compression algorithm: you're reducing size without eliminating the meaningful content. That requires a certain amount of judgement regarding what is meaningful in what's being summarized, in my view.

Open ##2380410
@tomstoneham@dair-community.social 2026-02-02 07:29

@DJGummikuh@mastodon.social LLMs are provably worse than that because they will completely ignore or actively change information which is unlikely. Classic example was a meeting notes app changing an extended leave request from a female member of staff (for special holiday) to a maternity leave request. We also see LLMs which aren't fine tuned for medical contexts not recording unusual patient symptoms or changing symptoms to align better with probable diagnosis.

Open ##2380413