Sunday story time. Today's ramble is on academia and using LLMs - sparked by commentary on bizgrindr (linkedin) about reviewing papers and academic work using LLMs.
If that is the path established career academics are taking, the entire academic endeavour is lost. Anyone promoting that stance should be considered unfit for the work,
Big words - hear me out, try to refrain from comments until /fin
At the core - science and specifically academia are about pursuing knowledge we don't know yet. Sometimes, if we're lucky, dancing at the edges of knowledge we don't even how to know yet!
The data don't exist. They're not anywhere. We don't know them, and we sometimes don't know how to know them or even if there's a way to know them. Serendipity leads as much as the structured work in a kind of chicken and egg problem.
/2
LLMs and automated systems, however, can only see the corpus of things that exist, and ways to connect them that already exist.
To paraphrase the classic MaricSteinian axiom** - we need new approaches to solve new problems.
LLMs are not able to do that. Statistically recombining snippets of analysis isn't novelty. It *approximates* the process that we *have done* in research.
That *have done* is important.
** I made that up, it's not how the axiom is normally described
/3
When we do a Research™️ - yes, understanding what we *have done* is necessary.
And then we break it. Or abandon it. Or do something else.
We don't grow knowledge by doing the same thing a lot of times.
[but isn't this the scientific method? repeat lots? Pay me a billion dollars to address that one it is soooo booooring. Or read the next sentence and *understand it*].
Repetition is how we confirm things. Not how we grow new knowledge.
...which is the job of academia and research.
Right?
/4
...in that framework, how can an LLM - a statistical reassembler of things we already know about - evaluate a research output?
You know the answer already: it cannot.
Ipso (w00t science word!), an academia using these systems is lost. And must be abandoned - recast into something new.
/fin
ps: everyone should congratulate me now for keeping that tight, there are *so many* side rambles to that one....
Oct 4, 2026, 12:41 AM@adamsteer as a journal editor, I can't express enough the joy I feel when I deal with a review that has spelling and grammar errors and is a bit chaotic and actually thoughtful. Human reviews are close to becoming the minority.