I wish they’d, dare I say, delved deeper into the bit about Nigerian English. My understanding is that’s where a lot of labeling happens. Now I wonder if that explains the idiosyncrasies not the size of the Nigerian English corpus online as the article suggests. If it was the latter I’d expect more “do the needful” and other Indian English idiosyncrasies.
You’ll notice, for instance, that A.I.s are absolutely obsessed with ghosts. [...] It knows that good writing involves subtlety: things that are said quietly or not at all, things that are halfway present and left for the reader to draw out themselves. So to reproduce the effect, it screams at the top of its voice about how absolutely everything in sight is shadowy, subtle and quiet. Good writing is complex. A tapestry is also complex, so A.I. tends to describe everything as a kind of highly elaborate textile. Everything that isn’t a ghost is usually woven.
Unfortunately, TFA is yet another article promising an answer to a "Why" question it never delivers. It documents the issue with many examples and thoroughly explores the variations yet never answers why current LLMs write as they do. It mentions some ideas in passing which might be causes but provides no justification or expert input.
Meanwhile, I'm pretty sure there are LLM researchers exploring this problem not only because it's annoying but because it's also interesting. If anyone knows of papers on this topic, I'd be interested...
I don't know papers but I figure the main reason why it's odd is humans have a picture or idea in their heads and want the reader get the picture or get the idea. LLMs don't and just string words together so it starts reading normally and then kind of wanders off.
Like for me the first paragraph reads normally up to "whisper woven from the algorithm" but then "a construct of code. A.I.-generated writing..." is just kind of muddying the picture with excess waffle.
I wish they’d, dare I say, delved deeper into the bit about Nigerian English. My understanding is that’s where a lot of labeling happens. Now I wonder if that explains the idiosyncrasies not the size of the Nigerian English corpus online as the article suggests. If it was the latter I’d expect more “do the needful” and other Indian English idiosyncrasies.
https://archive.ph/ntYnt
I like this theory.
Unfortunately, TFA is yet another article promising an answer to a "Why" question it never delivers. It documents the issue with many examples and thoroughly explores the variations yet never answers why current LLMs write as they do. It mentions some ideas in passing which might be causes but provides no justification or expert input.
Meanwhile, I'm pretty sure there are LLM researchers exploring this problem not only because it's annoying but because it's also interesting. If anyone knows of papers on this topic, I'd be interested...
I don't know papers but I figure the main reason why it's odd is humans have a picture or idea in their heads and want the reader get the picture or get the idea. LLMs don't and just string words together so it starts reading normally and then kind of wanders off.
Like for me the first paragraph reads normally up to "whisper woven from the algorithm" but then "a construct of code. A.I.-generated writing..." is just kind of muddying the picture with excess waffle.