Comment by tyg13
3 days ago
I don't really think that good writing and LLM writing looks all that similar. It's not always easy to spot (and maybe HN users aren't always doing a great job at it), but even the best LLM output tends to have an "LLM smell" to it that's hard to avoid.
Like, sure, LLM writing is almost always grammatically correct, spelled correctly, formatted correctly, etc., which tends to be true of good writing. But there's a certain style that it just can't get away from. It's not just the em-dashes, the semi-colons, or the bulleted lists. It's the short, punchy sentences, with few-to-no asides or digressions. Often using idiom, but only in a stale, trite, and homogenized manner. Real humans, are each different -- which lends a certain unpredictability to our writing, even if trying to write to a semi-formal standard, the way "good" writers often do -- but LLMs are all so painfully the same, and the output shows it.
I know the thing you are describing, but the real bitch is that you're actually just describing the lowest effort default outputs. The help-desk assistant persona.
Sometimes speedbumps that deter the lowest effort infractions are sufficient but I don't think this is that time.
On a per-prompt basis, or via a persistent system prompt or SKILL, or - god help us - via community-specific fine tuning, LLMs can convincingly affect insane variations in prose styling.
Seems like the ability to distinguish LLM versus 'good human' writing depends on the size of the writing sample you have to look at (assuming you think it can be done). And that HN-scale posts are unlikely to be a long enough for useful discernment.
Within a few years, LLMs will be indistinguishable from human text.
Think how easy it was to tell the differences a year or two ago. By 2030 there will be no way to ever tell.
The same is true of all video, and all generated content. The death of the Internet comes not from spam, or Facebook nonsense, but instead from the fact that soon?
You'll never know of you're interacting with a human or not.
Why like a post? Reply to it? Interact online? Why read a "news" story?
If I was X or Meta or Reddit, I would be looking at the end.
When will Teslas be self-driving again?
2 replies →
LLMs won’t destroy social media any more than it already is.
I don’t think I have ever had a meaningful human interaction with anyone on Twitter, Meta, or Reddit without already knowing them from somewhere else. Those sites are about interacting with information, not people. It’s purely transactional. Bots, spam, and bad actors are not new.
Meta has been a dumpster fire of spam and bots for over 15 years, the overwhelming majority of its existence.
Reddit has some pockets of meaningful interaction but you have to find them and the partitioned nature means that culture doesn’t spread across the site. It’s also full of bots and shills.
Nobody tells stories about meeting people on Twitter. At best it’s a microblog platform and at worst it’s X.
1 reply →
[dead]
AI driven web design has the same smell, it’s quite fascinating to see the different tells in different media. Then it’s also quite fascinating to see those same tells change and evolve over time.
Lol love the use of 'smell', that's a great way to characterise it.
It's not whether it "really" looks similar. It's what people think, most of the people, and most of the people are neither known for practising good writing nor consuming good writing.
LLMs have good writing in the same way that technical manuals can have good writing. It might all be correct, but it's usually not a good read.
Excuse me. I consider the writing within technical manuals strictly superior and meticulously written. It's fairly enjoyable to read what engineers/subject matter experts write about their own creations. Comparing those to LLM generated patronizing word vomit is a shame.
Depends on the technical manual and their culture. Red Hat had a culture of excellent writers, and their stuff is usually readable if not always enjoyable.
You're absolutely right!
(For those who have avoided reading AI writing, this is a trope referring to the tendency of some AI sometime to always agree with the user when corrected, I think? Or at least that’s as much as I have worked out, being one of those avoiders.)
Those sentence constructions that are "tells" were also learned from good writers though. But here, I'll let you be the judge. This was a comment I wrote 100% myself on reddit, which was both downvoted and I got multiple DMs referencing it and telling me to "stop posting this AI slop":
https://www.reddit.com/r/ExperiencedDevs/comments/1pyjkuf/i_...
Granted, it was in a thread about AI and maybe people were on edge, but I was still accused, which to be honest hurt a bit after the effort I put into writing it.
Interesting, that's one of the most AI-like comments I've read but it still feels human in a way that's hard to define. The headings, the punctuation, the word choices, the paragraph sizes all look GPT-approved. But there's just some catch in the flow, like inclusions in a diamond, that reads "natural" vs "synthetic".
I've been talking to Opus a lot lately though, and this could almost be something it wrote; it also has the tendency to write AI-ish looking blurbs that are missing the information-free pitter-patter that bloats older and lesser LLMs. People are going to hate me for saying it but sometimes it words things in a way that are actually a joy to read, which is not an experience I've had with other models. Which is to say, maybe what we hate about AI has less to do with the visual patterns and more to do with what we expect them to mean about the content.
But I think there will always be that feeling of: a human being took the effort to write this. No matter how informative or well written an AI article or comment is, it isn't something we instinctively want to respond to, the way we do when we know there is a person behind the words.
>But I think there will always be that feeling of: a human being took the effort to write this. No matter how informative or well written an AI article or comment is, it isn't something we instinctively want to respond to, the way we do when we know there is a person behind the words.
Over and over again, when reading comments from some folks who lionize the usage of LLM outputs, as well as other folks who demonize such usage, I'm reminded of this bit from Kurt Vonnegut's Cat's Cradle[0], specifically from the "Books of Bokonon"[1]:
And I wonder if, (myself included) those who demonize LLM usage are those who "came by their ignorance the hard way."
I'll admit that the analogy isn't great, but there is something to it IMNSHO. Mostly that many who distrust (and often rightly so) LLM outputs have a strong negative impression (perhaps not "murderous resentment," but similar) of those who use LLMs to spout off.
I suppose this is a bit tangential to the topic at hand, but if it gets anyone to read Cat's Cradle who hasn't already, I'll take the win.
[0] https://en.wikipedia.org/wiki/Cat's_Cradle
[1] https://www.cs.uni.edu/~wallingf/personal/bokonon.html
I think the comment you linked doesn't sound like AI at all, though. I do empathize with people worried about getting falsely accused of using AI in their writing, either hypothetically or in your case in actuality, but at the same time I kinda just think that's a skill issue on the part of the accusers.
This is very much a general "English reading skills" kind of test. A lot of people don't speak English as a first language, in which case I think it's entirely forgiveable. It's hard being attuned to things like writing style in a foreign language (I know from experience!). It's a pretty high level language skill, all things considered. And even among those who do speak English as a first language, there are many in this industry who don't have strong reading skills.
I do believe that personally my hit rate for calling out AI content is likely very high. Like many of us I've had the misfortune of reading more LLM output than is probably healthy for my brain.
One quick point:
>Those sentence constructions that are "tells" were also learned from good writers though.
I don't agree at all, I think the LLM style of writing is cribbed from like, LinkedIn and marketing slop. It's definitely not good writing.
As someone put ot before, humans use these little constructions maybe once or twice per article, not every single fucking paragraph.
This is a really interesting example because, to me, it reads as AI- or corpospeak-influenced human. I can't imagine anyone writing the text in the year 2000, but I believe you when you say you wrote it, and the actual information seems worth communicating.
It's the paragraph headings that look AI-ish. It seems to be rare for human commenters.
Notable exception being Stack Overflow-style answers, but I think those are more formal documentation and knowledge sharing / wiki pages than human comments. Human and more informal comments can be added as comments to answers.
I get that it's possibly contrary to the point if people are looking to truly have conversations here, but at least 99% of the time, I post a comment and never come back. I said what I had to say and don't particularly feel like getting sucked into an argument if someone disagrees, and frankly, if I'm wrong I think I'll realize it eventually anyway. I'm more likely to dig in my heels and ossify in a wrong position if someone shits on me and I immediately feel the need to defend myself. It can mesmerize you into believing things you might not have if it didn't hit your ego. I could be deluded but think I'm good at making arguments, but that at least means I'm good at making arguments that convince myself, which can be dangerous because you can convince yourself of things that are wrong. The upside is if anyone is out there accusing me of being an LLM, I don't even know so it can't insult me.
It is amusing to witness this happening to others when it's someone like you who is a semi-public figure who should probably be well known on Reddit of all places.
> It is amusing to witness this happening to others when it's someone like you who is a semi-public figure who should probably be well known on Reddit of all places.
One of our key tenants on reddit for a long time was "upvote the content, not the author". Which is why we made the usernames so small. It actually makes me happy when people judge the merit of what I write for what I said, not who I am.
But yes, it is sometimes tempting to say "do you know who I am??". :)
Nothing about that article screams AI slop to me. What a weird world.
I can’t help thinking how ironic it would be if your comment is from an llm
Poe's law strikes.
Parent's last paragraph was definitely an ironic portray of LLM writing! Notice the double-dash emdash.
Oh I saw that.
But was it really an ironic parody, or actually an llm?
1 reply →
LLM writing is like AI-generated photos in that you don't notice the good instances of LLM writing, i.e. you don't know your false negative rate.
I would say that you also don't know the false positive rate. The only person who truly knows is the one who wrote/generated the text. And they have every incentive to say it's not AI-generated, whether or not it truly is.
Personally, when I see the number of accusations thrown around, I very much suspect that the false positive rate is pretty high.
> I don't really think that good writing and LLM writing looks all that similar.
How do you know?
Confirmation bias; they don't know the LLM generated content they didn't recognize. They can't, because they didn't recognize it.
Pure AI slope os often extremely obvious, while for good AI writing that's indistinguishable from thoughtful human writing I'd say "Mission fucking accomplished"[0]
[0] https://xkcd.com/810/
> It's the short, punchy sentences, with few-to-no asides or digressions.
Uhh, isn't that how senior management in larger corporations communicates ...