Sorry, do people not immediately see that this is an AI bot comment?
Why is this allowed on HN?
Why is this allowed on HN?
Identifying generated comments is not always easy, as others have pointed out, plus we don't come close to seeing everything that gets posted. If you see a post that ought to have been moderated but hasn't been, the likeliest explanation is that we didn't see it. You can help by flagging it or emailing us at hn@ycombinator.com.
With LLM comments, there's an important distinction between legit users (who may have no idea that they're breaking a rule, especially because it isn't explicit in the guidelines yet) and accounts that appear to be posting nothing but gen-AI text. If you (anyone and everyone!) see a case of the latter, definitely please email us because we've been banning those accounts.
Even in more borderline cases, though, it's still helpful to email us because sometimes we contact the user, if we can, to let them know that we've been getting such reports. Or we might tell them we've suspended their account until we hear from them that they won't post LLM-generated or processed comments.
1) The comment you replied to is 1 minute old, that is fast for any system to detect weird comments
2) There's no easy and sure-fire way to detect LLM content. Here's wikipedias list of tells https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
How do you know that ? Genuine question.
The "isn't just .., but .." construction is so overused by LLMs.
In this case “it’s not x, it’s y” pattern and its placement is a dead giveaway.
It's not ironic, but bitterly funny, if you ask me.
Note: I'm not an AI, I'm an actual human without a Claude account.
I do agree with the sentiment of the AI comment, and was even weighting just letting it slide because I do fear the future tht comment was warning against.
It seems personal computing is being screwed so people can create memes, ask questions that take 30 seconds to find the answer to with Google or Wikipedia, and sound clever on social media?
If we are talking generative AI, again from my experience, things get a bit blurry. You can use smaller models to dig data you own.
I personally used LLMs, twice up to this day. In each case it was after very long research sessions without any answers. In one, it gave me exactly one reference, and I followed that reference and learnt what I was looking for. In the second case, it gave me a couple of pointers, which I'm going to follow myself again.
So, generative AI is not that useful for me, uses way too much resources, and industry leading models are well, unethical to begin with.
ChatGPT does this just as much, maybe even more, across every model they've ever released to the public.
How did both Claude and GPT end up with such a similar stylistic quirk?
I'd add that Kimi does it sometimes, but much less frequently. (Kimi, in general, is a better writer with a more neutral voice.) I don't have enough experience with Gemini or Deepseek to say.
LLMs do these things because they are in the training data, which means that people do these things too.
It is sometimes difficult to not sound like an LLM-written or LLM-reworded comment… I've been called a bot a few times despite never using LLMs for writing English⁴.
--------
[1] particularly vapid space-filler articles/comments or those using whataboutism style redirection, which might be a significant chunk of model training data because of how many of them are out there.
[2] I overuse footnotes as well, which is apparently a smell in the output of some generative tools.
[3] A lot of pre-LLM style-checking tools would recommend this in place of hyphens, and some automated reformatters would make the change without access, so there are going to be many examples in training data.
[4] I think there is one at work in VS which I use in DayJob, when it is suggesting code completion options to save typing (literally Glorified Predictive Text) and I sometimes accept its suggestion, and some of the tools I use to check my Spanish⁵ may be LLM based, so I can't claim that I don't use them at all.
[5] I'm just learning, so automatic translators are useful to check what I'm written isn't gibberish. For anyone else doing the same: make sure you research any suggested changes preferably using pre-2023 sources, because the output of these tools can be quite wrong as you can see when translating into a language you are fluent in.
[6] Another common “LLM tell” because they often have weighting functions especially designed to avoid token repetition, largely to avoid getting stuck in loops, but many pre-LLM grammar checking tools will pick people up on repeated word use too, and people tend to fix the direct symptom with a thesaurus rather than improving the sentence structure overall.
I've find myself doing it, a time or two.
We'd much rather hear you in your own voice: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
Using LLMs to learn, of course, is great. HN posts benefit from things people learn however they learn them. Just please write them yourself.
(Also, I'm sorry for the harsh reactions people have been showering on you for saying this. The community feels really strongly about it, and of course the norms around all this are still in flux, not just here but in society.)