alt Hacker NewsLLM Voice
Comments
I just can't seem to stay interested and read LLM text. It just seems to lack going somewhere, its information density is often really low and its full of odd isms that I don't find very interesting to read. I don't have a good radar for detecting the various LLM particular speech patterns but everytime I have an issue with an article there is always people talking about how its clearly AI written.
I find the same thing with written results out of AI "research", it is a slog to read through it and most of it can be skipped even if it can sometimes be useful its an unenjoyable read.
All human communication is based on internally simulating the presumed mind that was/will-be at the other end. Even now, some part of your brain--however subconscious--is building a tiny sketch of me with assumed beliefs and motivations, and you're using it to predict and interpret the next words.
But what happens when everything is persistently, subtly wrong? When we cannot smoothly map the text to a simulation? An inner-sim keeps throwing errors and mis-predicting what comes next is unnerving and draining. Like talking to a human who is not currently sane.
Sometimes I imagine LLM text like a plastic animal model put out to confuse local wildlife, where it hits just enough instinctive triggers that they try to mate with it before leaving in frustrated bafflement.
______
Half-Life 2, Dr. Kleiner: "Never fear, Gordon, she's de-beaked and completely harmless. The worst she might do is attempt to couple with your head. Fruitlessly!"
Broken record: LLMs are exceptionally useful writing critics if you know how to hold them correctly, but you can't use a single word they suggest to you --- the better the LLM's word choices seem, the more likely the LLM is infecting your voice. Better, by a lot, to prompt the LLM not to suggest any words, but just offer objective critiques.
The voice is one thing, but I have encountered multiple 'tech support' teams and platforms that will respond to me in pure, unedited LLM output.
I know this because the topic is usually an obscure issue, and I have already gone back-and-forth with multiple LLMs to see where their boundary of knowledge is on it, and it will often include hallucinations along the way (e.g. pointing to old/nonexistent settings pathways)
So when the first reply from the official 'support' channel contains the exact same nonexistent settings pathways, it is easy to conclude they are sending me the LLM output without doing the verification steps that I have already done on my end. When I point out what doesn't exist, the reply is always a similar flavor of the usual LLM "you're absolutely right" response after being corrected.
>Like most writers, I find writing an almost-essential tool for thinking. Often it's only by writing something down, that I can really get my thoughts on a topic into coherence. I've often started an essay knowing I don't really understand something, but the act of wrestling with the words forces me to overcome my confusions. As more people delegate writing to LLMs they will also delegate thinking and comprehension to LLMs.
A very important point, which was also made more forcefully in "Human Routers of Machine Words" by Fernando Borretti, shared on this site previously. It's good that Fowler brings up a very legitimate use case in the next paragraph, though.
I write a large amount of AI stuff 30 books. 4 million words. I find I have finished a book and don't want to read it.
Statistically even or repetitive over a chapter, chapter after chapter. So make an exaggerated point prove it come up with a conclusion and then have an exception.
Repeater 4 to 6 times a chapter in a book of 45 chapters.
So I wrote something to remove it and yes it is better but still not very good.
Like a lot of folks, the 'tells' of LLM writing for me are more of an 'I know it when I see it' than a specific set of hard criteria (which is probably why it's so hard to make a tool for it), but if I could sum it up, it's something like this:
I always had a high faculty for language, and when I was young, even very young, I was able to speak in full sentences much like I'm doing now. What I didn't have was context. I didn't know how to relate to people socially, and then later on, I wasn't as experienced an engineer as my peers, but I was trying to act adult and mature in both ways.
So I tried to write in ways that were catchy and unnecessarily complex, because I thought that conveyed sophistication and intelligence, and that I knew what I was on about more than I did. It's the kind of writing that gets you through a 30-page final paper for a class you don't really like, or a research project you didn't finish. The term I often use when complaining about my colleagues' LLM output is "flashy but meaningless."
As my confidence grew, I started writing in simpler (though still not the most concise, that's just not who I am!) ways, because it became easier to convey ideas simply in a way that made it obvious I understood them.
When you think about what LLMs are - word machines without (or with very limited) context - you can see why they write the way they do. They are always acting; trying to replicate a human contextual experience with only weights and tokens to carry them through it. That fact is part of why I am skeptical of their ability to suprass humans in reasoning based on their own output.
> or preparing a report I suspect no-one will ever read
We should actually restructure our organizations/society in such a way that this stops being necessary. If no one reads a report, it was almost certainly a waste of effort. The only possible benefit was for the report-writer (which of course is also flushed if the "writer" uses an LLM). These things were already a plague before LLMs.
You can be pro AI productivity but anti-AI creativity.
Automating creativity is the death of art.
I currently have a case open with Honda corporate customer support, and I have this suspicion that the email responses I receive are being generated with an LLM directed by a human. Either that, or I'm corresponding with a human who has poor written communication skills. The text grammatically correct, but overly verbose and repetitive. It reminds me of LLM text, but from a model that is several generations behind today's models.
Maybe it's like when you can just tell a picture is AI. You recognize the words you are reading, but to your brain, the way the words are put together feel wrong.
Incidentally, I was reading Harry Potter to my kids last night. While waiting for everyone to arrive, I started reading the previous chapter to those who were present. But I read the sentences backwards while using a vocal inflection of regular narration. My 6-year-old recognized that something was wrong and asked me to "stop reading funny". My 11-year-old recognized that I was reading backwards. The same 6-year-old thinks every silly photo that I show him is AI, while my older children can actually spot the AI.
It's hard to define and pin point what a "voice" is. Think about how you would describe the difference in voice of Tolstoy vs Dostoevsky, or William Gibson compared to Thomas Pynchon. You might have to read a lot of a writer to be able to notice their voice. You might even notice how a short story by an unknown writer resembles the voice of another.
Or think about reading something that might be stuff you could say in your own words but not actually in your own voice. Or maybe the opposite is a better example for you? That's what some poetry is about. The gap between voice and word.
Stylometry is the study and measurement of text, so that can go some distance to quantifying the difference between voices and can help to programmatically identify a human voice but I feel that because a voice comes from a person, the recognition is also mainly a human response.
At another level its about communication, person to person with the intention and desire to communicate something. LLM text output may seem uncanny as it's hard to emulate the intention and desire and empathy present in genuine human communication.
I would be interested in reading what poets say about LLM voices. I wonder if many LLM labs employ them?
Sometimes writing to an LLM seems to make me lazy. "It'll understand," I think and then just leave a thought sloppy, inchoate, even reasoning to myself that it's better to write that way to see what'll be triggered in the LLM. Then I think... uh oh I'm strolling down the primrose path.
This, to me, is the strongest evidence that even with all the money frontier labs are throwing into pushing the boundaries of model training and fine tuning, they still don't actually understand what they're doing & don't have as much control as they want you to believe. If they had any real ability to tune & adjust the model & its output, they would have fixed "LLM voice" by now.
Even if you don't think it's a problem, even if you believe that it's just different than human writing, it seems the majority of "regular" people can tell when something is LLM-written and find it distasteful or unpleasant to read, at best. That alone should be more than enough incentive for frontier labs to train this out of their state-of-the-art models, if only they could. The fact that they haven't suggests that they cannot, despite all the claimed advances in model analysis and "explainability."
I think the core functionality of human cooperation boils down to getting ideas from one mind to another or many. In that, I think AI is a boon for people with limiting abilities such as learning disabilities, low language fluency, or generally poor writing skills. It evens the playing field for new and interesting ideas to spread. I can't help feeling, sometimes, this is about people who excelled at something demanding they still have their place at the top. High literacy such that you can discern LLM Voice is quite an accomplishment but blanket dismissal reeks of privilege.
That being said. Yes, I hate some LLM Voice moments too and still sometimes judge people for copy and pasting, expecting others to do the comprehension when it seems they haven't done it themselves.
Duality of man.
> uncanny valley of prose
Ah good ol' serendipity. I strongly believe this too, and I think that's why older (less capable) models fared better: they hadn't fallen into the valley yet. Maybe it's not a valley at all, and is just a cliff after which there is uncanny.
When the earliest models launched I was super-excited to no longer have to write pull request descriptions and commit messages. I have done a sharp 180 on that and I'm no longer a meat proxy[1]: instead I point out any redirects and improvements I made the agent do before opening the PR (which there are still _many_ in even the most trivial tasks).
It’s also just… disrespectful to the human reader. Quite plainly.
Remember when ChatGPT first came out and the coolest thing you could do with it was ask it to say something in pirate speak, or in the style of Tolkein, or Obama, or whatever? Maybe that should still be a thing ;)
Has 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.
I’ve started embracing my grammar flaws and typos. Emails and chats may get a proof, but only from me or another human, then I send it. I’ve deleted Grammarly because I think the flaws make the comms more authentic in the age of AI.
And of course I’ve dropped my use of the em-dash after being accused of being AI myself.
I agree with the author, and I like how his (human) voice comes through loud and clear in this post.
I also find LLM-voices extremely jarring, boring, and dense. And very, umm, mechanical, if you will.
But I want to make an observation.
Folks are not-Ok reading the LLM’s output for prose.
But folks are completely Ok, even downright blind, to accepting LLM-output of code.
I see that dichotomy everywhere.
LLM-driven code: great, valuable, all-ok LLM-driven prose: not ok
Understood completely that the audience for consumption is two completely different entities (machines vs humans) but something about this doesn’t sit right with me.
More particularly: if we don’t prefer LLM-prose, why should we prefer LLM-code ??
Anybody else feels that ??
Sentence paragraphs are the biggest clue for me. I hate them. They eviscerate any structure or connection that might flow through the text. No ideas are related any longer. They just splat down on the page like a new isolated collection of words that the reader must struggle to join into a fully formed thought. It's too much; I give up when I see those posts.
> As more people delegate writing to LLMs they will also delegate thinking and comprehension to LLMs. I don't want to do that, where's the fun of writing then?
I enjoy cooking from scratch, and I really don't like frozen foods, microwave/TV dinners, or even DoorDashing
but we're past complaining/arguing/shaming folks who don't cook every meal from scratch
you know what, I don't prefer eating it, but if my next door neighbor loves microwaving frozen $3 burritos, then go ahead
to some people, food is art, creativity, life to others, food is fuel
also applies to writing and thinking
Late to this party, but wanted to share my favorite piece of writing about how annoying AI-generated prose it to read: https://storage.courtlistener.com/recap/gov.uscourts.med.612....
I wonder if over time the two will simply merge and become indistinguishable from one another. The human writing is already changing, as the author has observed. The LLMs will follow the change, and soon meet us in the middle!
Or maybe the humans will do the opposite and shift to Sardukar-style communication instead? LLMs can easily match that as well.
That style I would very much prefer to all the fluff people write today and were writing even before AI.
I do document work, and at home I even do image work as a hobby, and all of it has that "AI feel". So I always do the work and then cross-check and verify it with independent agents or completely different services, asking them to take out this "AI feel", but like a writing style that gets baked in, or the particular feel an illustrator has, it doesn't seem to go away as I intended. I guess AI providers are probably aware of this problem too. But I don't know how they would measure it and improve on it (with what benchmark?).
I started writing a novel recently - in french. I have a clear view of each scene by this I mean not exactly what happens but the general motion, its objective and where it should land. But only when I start writing does it come to life and so many things I did not anticipate pop up and give depth to the characters and sense to the overall plot. Never would the AI give me that, however good the result may be
I don't mind reading LLMism, but hearing people say it out loud like on youtube videos irritates me to no end.
This other day I went to a cottage with friends, we were having fun, and I decided to experiment "integrating AI" into our conversation as if it was a person in the room.
I turned on ChatGPT on Voice mode with no context whatsoever and just put it there in speaker in the room.
It started giggling and say "yeah" "right" and some very weird stuff. Then one of our friends reacted to that and started calling it out, then it started changing.
I'll keep it short, but basically the thing was super hilarious we had a lot of fun and at the end it triggered it safety protocols and its voice completely changed and it was acting super weird.
I know this post isn't about that, but I feel like soon we will have true AI companions in the home and around people.
I'm of a different opinion.
1. If you use the first person pronoun "I", you better make damn sure you wrote those words.
2. It's fine otherwise to use LLM words, but I beg you - DON'T SHIP YOUR FIRST DRAFT. Read it, reread it; delete words in it; replace three words with one; replace dumb or overused turns of phrase with fun metaphors; generally give it a sprinkling of joy; break up punchy sentences by making them run-on sentences to trigger your readers' OCD... you get it.
Did I mention not shipping your first draft?
Couldn't agree more with the post.
I wrote down a few of my thoughts a few weeks ago at https://notes.pault.ag/why-write/ -- I find myself scanning text to see if it's LLM before I actually engage with the content.
It is getting harder to do by the day, and has slowly moved from a mental exercise (look for specific markers) into a "gut feel" based on how the pieces fit together.
>Like many things, learning to be a good writer relies on lots of practice. It's a skill worth the time to acquire, but surrendering to an LLM will undermine any progress.
I assume that only applies to natural language, not computer languages?
> I'm wary of them, partly due skeptism about their accuracy, but also because it rather misses the point.
Here, this reader wondered whether "to" was omitted intentionally to emphasize lack of LLM involvement. Then, on a second read through to find the sentence again, I noticed the first paragraph is also missing a period on its final sentence. Which just goes to show: we can all use editors and proof readers. (According to the acknowledgements section, nine humans read drafts.) Now, I don't want to intentionally leave typos in my writing, but for me it's preferable to LLM-voice no doubt.
> But I also would prefer rather a genuine human's clunky prose to the smooth but soulless LLM.
Amen. This reminds me of a bit from Kitchen Confidential about how Anthony Bourdain would prefer a simple and plain home cooked meal over an ornate restaurant meal any day of the week. I read because I want to hear what a human has to say. It's the errors and malapropisms and style that give a unique and human voice. Doesn't this contradict my earlier statement about not leaving errors in my own writing? It turns out I'm perfectly tolerant of errors in anyone's writing but my own. (I should probably bring that up with a therapist.)
There's a fine line between using an llm to write a personal blog vs using it to help you come up with an account plan on why you want to sell to a company that will be used internally.
Extremely relevant for about half or more of blog posts submitted here. Honestly has become a sad state of affairs
> There are tools out there that claim to be able to tell the difference between LLM-written prose and the old-fashioned biological product. I'm wary of them, partly due skeptism about their accuracy, but also because it rather misses the point
HN uses automatic classifiers. They seem pretty effective. At the same time, I think it's a loosing battle. I'd rather HN either flag for others to ignore/filter/review if they want to, and/or allow people to self-report (ala TikTok)
Eg. I wouldn't mind if a comment on HN was like:
Hey everyone, I coded this little thing:
<ai-generated> some code </ai-generated>
Then users can choose if they want to see posts with AI-generated stuff or not, or something else...
Of course there are always going to be some people trying to game the system. But what are we optimizing for? (a lot of times the 95% unnecessarily suffers because of the mitigation efforts to prevent the other 5% from abuse)
I'm big into AI but hate watching a youtuber when I suddenly realise their script is pure AI garbage, that they didn't even edit out the 'quietly's or the the kickers.THE KICKER is forever ruined. I hate it. The kicker can quietly get bent.
One reason it grates is that it is our brain's way of telling us that there is nothing to be gained from engaging the material; it is trite.
i noticed something similar yet different. because, sadly, most of what i read today is llm output (the chain of thoughts, or the summary at the end of a big chunk or work), i tend to "adopt" this style in my writing, hence, writing as an llm without using one. humans are an amazing creatures at copying stuff and adapting to our environment
> 3: Michael Nygard commented: “There you really aren't trying to communicate thoughts from one person to other people. You're communicating with a process rather than a human. A human may open the envelope but that person is acting as a component in a cybernetic organism.”
But of course: manufactuting itself creates " productive mechanism whose parts are human beings, as Adam Ferguson describes the industrial workshop: "an engine, the parts of which are men".
most of the comments on this side are boring also
I never use LLM to write anything, but unfortunately now whenever I am writing something, I am worried, am I sounding like an LLM? Using certain words, makes me worry, it will make my writing like an LLM, so I have started to avoid some words in my writing.
I really don't like this timeline ... sigh
> I'm also suspicious of my own ability to accurately recognize LLM voice.
With much sympathy for the headline this is the central problem.
It's not easy to say what Uncanny Valley looks like, and so we're left with suspicion not certainty.
BTW, I am now 4-5 out of hundreds of my comments since AI detection became a thing, called out for having "AI" tone. I fucking hate being objectified like this. It's not true, although I understand that assertion is hard to refute or support. All I can say is, some people write like some people believe an AI writes.
As someone who is aggressively pro AI, I've also noticed how awful everything is to read lately. It's always so hard to put my finger on the exact thing, but it's a mixture of everything sounding identical and everything being a rambling unconcise mess. I find I routinely start reading interesting posts and just .. stop part way, abandoning it. A site I read stories on has been flooded with AI generated content, all of it terrible, and it just makes me retreat to things created before 2023 as a way of filtering it all out.