alt Hacker NewsHas anyone done any work on what it is that makes LLM Voice hard? Something I've noticed, and have seen other people comment in passing is that reading LLM output is "exhausting". It's often attributed to trying to detect the machine, or filter out BS or word density or some other explanation, but it feels like something more than just that. Like - there is something about LLM output that just doesn't "fit" our minds correctly, even though words are in the right places.
Surely someone with a neuro/psych background has picked this up and developed an evidence-backed explanation.
Replies
I read a lot of old material from 1-1900AD where I can pick upon the author's 'voice'. There are subtle clues, like if they had a regional accent, where they use a specific word or type of cadence.
AI does the same with responses where the cadence is some type of formula that becomes identifiable.
Each of the big models has their own but an example of the cadence is
A) Agreement or Acknowledgement with the User
B) Specific detail gets a paragraph of detail related to the previous response
C) Overview of the conversation's current state
D) Prompting or Direction for this conversation to go
E) Landing on a beat. Wrap up the vibe. Recap with a zinger.
I cannot unsee it and a few of the models are annoying. I have attempted to build out the Agents and context to avoid the 'last beat one liner' without success because after 2-3 more turns, I get "That challenge was accepted and you stepped up. You're the winner this round".
There is this article on this topic, which I found quite interesting: https://arxiv.org/pdf/2610.29162
Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation. [...] We propose and test a hypothesis as to why: while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for information density rather than successful communication.