How LLMs Distort Our Written Language
41 points by vesto
41 points by vesto
This is deeply disturbing to me. The example given in which the LLM changes the argument about the self-driving car is especially upsetting.
It is not shocking to me at all that the LLMs take neutral stances: as far as I understand mainstream LLMs, this is essentially a core goal of them as products. That is, they are only supposed to argue for "known" truths and in support of the user, and otherwise equivocate and take the middle ground.
It is mind-boggling to me that people reach for LLMs in order to write or edit anything of value.
It is mind-boggling to me that people reach for LLMs in order to write or edit anything of value.
Even in the charitable interpretation where someone writes a completely human-made first draft and then hands it to an LLM to "check for issues", how do you not then examine the changes made and see that the resulting text has a different meaning than what you submitted?
"Disturbing" is exactly right.
While I don't use AI and largely agree with you, the OP demonstrates that LLM-influenced writing (as opposed to writing in which LLMs are heavily leaned on) are semantically fairly aligned with LLM-free writing. The paper argues that only heavy LLM usage (which is quite common!) significantly impacts the semantics of writing.
they are only supposed to argue for "known" truths and in support of the user, and otherwise equivocate and take the middle ground
IIUC, Musk's Grok is/was intentionally biased.
That is, they are only supposed to argue for "known" truths and in support of the user, and otherwise equivocate and take the middle ground.
This is good professional practice for a secretary or copy editor.
I don't understand why you think it is bad when LLMs are trained in this direction.
I saw this a lot when I was trying to use Claude as a copy editor. It took multiple iterations on the prompt to get it to pay attention to spelling, grammar, and punctuation only. I suspect the tendency to shift meaning has to do with the way embeddings work.
The frequency graph is jaw-dropping and looks, honestly, pretty much exactly like I expected it would look.
Consider this a gift: those things on the left: those are words that are now powerful. Those works on the right, those are words that are now (increasingly) meaningless
Which graph? I didn’t see anything like that on the page
I was confused at first, too. I think they mean the word cloud near the top, above the Executive Summary, with labels “Unique Words from Human Edits” and “Unique Words from AI Edits (GPT-5-Mini)”. In the paper, it’s Figure 3.
Why should we care?
As LLMs are integrated into society, these subtle changes in meaning could fundamentally alter politics, culture, science, and even the way we communicate with our friends and family.
The agency and responsibility of users seems curiously absent in much of the discourse surrounding AI / LLMs.
Assuming the user is an adult, they actively choose to use an LLM. They can decide whether or not to use the outputs, and in what way.
If AI does "fundamentally alter politics, culture, science, and even the way we communicate with our friends and family," then it will be because people chose to do it and AI facilitated that choice.
Lack of opinion or preference on the part of the user doesn't obviate the fact that a choice was made.
I sometimes use LLMs to catch grammatical issues, inconsistencies, places where I overuse a specific word or phrase, and occasionally to offer suggestions to help me simplify something. I read its suggestions, and implement them myself if I like them. I don't let it edit, and I rarely accept any suggestions that change my meaning.
I'm a lot less picky about the code it writes though. I complain if it hardcodes things, I complain if it names things poorly, I complain about code duplication, I complain if it tries to disable a warning, I make it pass lint checks and run code formatters. I often inspect its work, but rarely complain about its algorithms. I try to work a couple of hours every day without it, because I like it, but I have to admit, I also like bossing it around.
Read this along with the new paper on subliminal learning. Imagine an LLM trained by a particular entity that wants to affect how it is perceived in the world, and is made available for free.