[precompute]

I have a lot of experience with "LLM voice" as well, and none of that sounds even remotely LLM-written.

The smoking gun for LLM-written text is when the text is a "linked list". It can only ever directly reference the previous thing. That's not the case here. And the latest Hunger Games book isn't yet another amazon-published slopfest. It's been through a couple of rounds of editing, at the very least.

I'm not saying RedditOP is completely off-kilter. There might be something to what he's saying. Maybe Suzanne Collins (the author of the book) has been consuming a lot of LLM-generated content. Or maybe she's just ahead of the curve and writing in a style that's likely to catch fire (no pun intended) [1].

[1]: Yes, I wrote this myself! And the entire reply!

[lloydatkinson]

> none of that sounds even remotely LLM-written.

Did we read the same extracts? The nonsensical actions and movements of the lovers in the entire train scene? The obnoxious call and response structure? The absurd comparison between a grandmother skin and a spiders web because "silk"?

> It can only ever directly reference the previous thing. That's not the case here. And the latest Hunger Games book isn't yet another amazon-published slopfest. It's been through a couple of rounds of editing, at the very least.

I don't agree with your assessment here, "the last thing" can be literally anything the user prompts. Are you suggesting that because none of the previous books in the series are written by AI, that's somehow an argument that the latest in the series can't be?

[precompute]

Okay, first, I haven't read that book. I'm going off the cherrypicked examples in the OP.

If a LLM was indeed used, the output was likely massaged to a degree that wouldn't be immediately obvious.

Now, my 2c: Writing is sometimes atrocious and sometimes authors jam in stupid things to maintain flow. If a person could make the connection between "Silk", "Spider", "weaving" and "grandmother" then another person could as well (even when one is "verifying" and another "proving"). And using those properly and in context and succinctly is far beyond most LLMs, and would require a fair amount of gambling, which would be out of character for someone whose writing prowess has been verified pre-LLMs.

As for what I mean by "directly reference the previous thing": LLMs can jam in well-known (to them) phrases and sentences and structures onto an idea/request. However, they are unable to loop upon the particulars of that idea/request in a coherent manner, which leads to slopification at large output sizes, and shows us the ceiling of the quality of writing a LLM can output.

[orwin]

I use AI for world and character-building in the RPG i DM, and this is for sure something an AI would write. AI i use professionally do like linked list, but when i ask it without a MCP, to "write" a character and its background, catchphrases and what he would do in 3-4 different situation, you'll get paragraph that suspiciously look like the first example.

[dns_snek]

LLMs can change their "voice" with a simple instruction. If detecting LLM-generated text was as easy as you think it is then services in this space wouldn't suffer any false-positives or any false-negatives.

[precompute]

Oh I don't think it's easy or that one can be completely sure about it. And in my experience, LLMs aren't very good at changing their voice. If you have any examples to the contrary, I'd like to see em.

[dns_snek]

I don't have examples, I'm asking you to provide evidence which supports your extraordinary claim.

You claimed, with a high level of confidence, that the text isn't written by AI because it lacks obvious "tells" which you believe to be present in any LLM generated text. But if the absence of these "tells" reliably indicated human writing then LLM detectors would have false negative rate of approximately 0%, do they?

[precompute]

It's not necessary for the tell to be easily computable. It's not something that's true everywhere, but it has been true more often than not when I've tried to use LLMs. And I haven't seen a LLM-generated piece of writing that's sufficiently long and complicated enough to rule this tell out.