The Voynich Ninja

Ah damn, this again. I suppose everyone know this feeling that when you're reading something, you think "this is too technical for me, other people probably understand this much better".

One of the major issues with AI in scientific writing is that it takes a huge investment by the reader to distinguish between this phenomenon and AI talk. Do I have a hard time reading this because it's highly specialized, or because it's not entirely written by people? This explains a lot of what we have seen (and will see) when slop gets past some filter it shouldn't have.

Anyway, I am unable to judge to what extent this publication is affected by it. I do believe in the principle that scientific writing should aim for clarity of language in all cases though.

I think I'll split off the irregular spaces question I had.

(22-08-2026, 06:49 PM)Koen G Wrote: You are not allowed to view links. Register or Login to view.One of the major issues with AI in scientific writing is that it takes a huge investment by the reader to distinguish between this phenomenon and AI talk. 

The human editing perceptibly falls off as you get toward the end of the paper, and throughout the sudden Thor hammerstrikes of rhetoric combined with "earn" crap phrasing is dirty. Limitations section isnt human written. My take on a lot of this is if you can remove the slogans and bumper sticker phrasing AI introduced there's actually some good nuggets here. The entire appendix has been abandoned by humans, it is iRobot. Its crap like "the graded boundaries are where an account of the text has still to earn.." and tjhen "joint profile is only useful if it discriminates…" blsahblahblah

But the findings are good stuff, they just should've used Fiver and paid a law or science journal editor to fix it up before uploading.

TLDR; genuinely good stuff stinkified by AI phrasing

Disclaimer: I'm a US lawyer and we're 1+ year into being slammed by AI slop and hallucinations in court filings so we're all PTSD on chatgpt and claude phrasing trauma.

JB

The other AI fingerprint is a number of the sections (and you are absolutely right the appendices are even worse about this) are way, way too short. There is not nearly enough explanation of their methodology nor justification for it, and the conclusions are very much from the "the numbers explain themselves" school of thinking. While these can be problems with human writing, especially when the human comes out of a field where the expectations are practical rather than academic, that's not my sense here

As someone who hates AI writing for it's own sake, the "must earn" tic is most irritating to me for being a reasonable point---both in context and in more generally for those of us modelling these problems