Proposal: add "AI generated" as a flag reason
294 points by danderson
294 points by danderson
IMO, the feed is being DoSed by LLM-authored text and vibecoded software. Anecdotally, the community's immune system is currently rejecting these in two ways: adding the vibecoding tag to anything that's been touched at all by an LLM, or flagging posts as spam.
I suggest that both these immune responses are problematic in their own way, and that giving us a way to flag specifically LLM generated content would be better.
Per the guidance on the lobsters code repo, posting this as a meta first to gauge sentiment. If this is something we'd collectively like, I would be delighted to make the implementation happen.
To define what I mean by AI generated: text or code that was substantially not authored by a human mind trying to communicate an idea, but rather by having an LLM expand a prompt into a larger artifact. There is an element of Stewart's "I know it when I see it" in this definition, but https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing is one laundry list of hallmarks of LLM output.
Speaking personally, I think flagging as spam is the correct response. Such content is a denial of service attack on our collective attention. I find it deeply frustrating to get partway into a submission before seeing the LLM signs, and realizing that I've likely already spent more time reading the slop than the author spent writing it. This type of content is no better than what was on the content farms and lazy clout farming blogs of old, LLMs have merely made it much easier to generate such content in volume.
However, flagging as spam seems to cause confusion. It seems to me that the confusion stems from people who haven't learned to recognize LLM text, and express a confusion that boils down to "this doesn't look like commercial/ad type spam, why the flags?" Having a separate flag reason that conveys "hey this is AI slop" would address this confusion. Incidentally it'd also help folks learn what LLM text looks like, which I think is an immensely important skill to pick up.
Likewise, flagging everything that's had an LLM near it as vibecoding seems to frustrate people. I personally think we'd be better off with all the vibecoding posts deleted from the site, but let me try to steelman. I think the contention is that there is interesting, high-quality stuff out there concerning the application of LLMs to software development, and that the tag should be concerned with those rather than acting as a dumping ground for anything that an LLM has touched. If this is correct, then having an explicit flag for removing low-value LLM output should benefit both the people who seek out posts tagged vibecoding, and those who hide the tag.
(I'm skipping over the whole "rename the tag" discussion, which has been rehashed several times already in recent meta posts. I think it's orthogonal to this proposal, and I don't think this proposal provides any new data for that question.)
I remember seeing post deletions in the moderation logs with the reason (paraphrased) "This is 'I vibecoded something', which is off-topic". I argue that this is a special case of the general rule I'm proposing: content whose human involvement was mostly prompting an LLM and (maybe) skimming its output is not a welcome contribution to this site. And I would like a way to explicitly call out and flag such vibed content, code or not.
As you mentioned, maybe AI slop is even better than AI generated (to prevent it being used when someone posts an interesting vibecoded story).
actually, why not just "slop"?
human slop existed before ai(really, half of medium and dev.to), and i dont see any reason to tolerate that either.
this. just slop. what matters is the quality of the final product, which one could theorize is the best indicator of the amount of human effort that went into it.
what should matter is either just the quality and newsworthiness, or how human effort was filtered through tools, not what those tools actually are.
I've been lurking for a long time, but decided to bum an account here in part because something upward of 20% of stories on HN are AI generated and I find it exhausting.
I'll say one thing: I think the distinction matters a bit. Writing used to take effort, so articles existed for a reason. Sometimes, that reason was self-promo / ad clicks, but there was a reason.
Now, AI reduced the cost of producing content to nearly zero: "Gemini, write a provocative piece about why vibecoding is great / terrible". So, it's possible to conjure an infinite stream of reasonably articulate opinion pieces just for the sake of it. It's both a quality problem (is there any discernible reason for this article to exist?) and a volume problem. There are only so many hours in a day, and even if the quality is comparable, I prefer to spend them reading human work product.
I think that if you don't explicitly tip the scales in favor of human output, you're mostly gonna be reading LLM output in a year or two. I don't have a good answer here, but my point is that it's not just about quality. In fact, if your prompting is half-decent, the quality can be half-decent too.
Yeah I recently joined for the same reason.
I don't know exactly what the heck has happened to HN last few months, but the quality of the submissions have taken a massive nosedive, and the comments are in many cases much more banal than what I've come to expect from the site, closer to what I'd expect off Reddit.
Very much getting a dead sea vibe. The valuable participants are evaporating, leaving only the ones that don't mind the slop.
Would be very interesting to have a better grasp of what is driving this community rot, so that it can be prevented from spreading here. All we're sitting with now is a vague sense that LLMs are driving it, but how? Why has it seemingly accelerated recently? Has it?
The way I see it, HN is 50% a replacement for Slashdot and 50% the gathering place for the SF Bay Area hustle culture. The latter flares up every now and then, most recently during the cryptocurrency / NFT boom.
Right now, I think many folks are convinced that we reached a "business singularity" and you're not building with AI and plugging AI into your day-to-day life, you're gonna be left behind. I don't think this is even necessarily wrong - there's a lot of cool stuff you can do with automated natural text comprehension alone - but I don't need to read ten LLM-generated think pieces about AGI every day.
I would even cut it down to "slop" so there’s no arguing about the level of AI involved.
If it reads like a LinkedIn post, it should probably be flagged as "slop"
I think this puts the focus where it should be, that is, on whether there is substance and thought behind the work, and not whether all participants can pass the Voight-Kampff test.
There's already people who are afraid to use certain words or turns of phrase because of that sort of misplaced sentiment. That's not a good vibe.
That would be my preference, yeah. It's short and sweet to match the other flag labels. I suppose it's also a better middle ground for the community, in that it aims specifically at low-effort, low-signal content. That leaves open the more contentious question of whether any other kind of AI-generated content exists :)
I vote in favour of this suggestion.
Personally I am interested in the conversation around AI (and try to be less angry about it than I have in the past), but I filter the vibecoding tag because I don’t want to read slop. Being able to distinguish genuine thoughtful arguments about AI from generated ”content” would be helpful.
In order to sidestep any arguments around what counts as AI generated, or the "who cares how it was written as long as it's good" comments, what I suggest is to add a "low effort" or "low quality" flag.
AI-written / AI-edited posts usually fall under this category, as do many posts about entirely AI-coded projects. I wouldn't want any posts flagged with the AI-generated reason to fall into these meta-arguments all the time.
There's the question of: if it's low quality, why are people posting and upvoting these submissions in the first place? I think, like you mentioned, not everyone is aware that what they're reading is AI-generated, or perhaps they only quickly skimmed the post before upvoting. Perhaps they're not even clicking through! (I know we're all guilty of upvoting a submission before carefully reading the linked post from time to time)
There's the question of: if it's low quality, why are people posting and upvoting these submissions in the first place?
You don't need so many people enthusiastic about a topic to flood the front page. Lobsters has 20k users. There are right now only 7 stories in the front page with >50 votes. (BTW, only one of them has the vibecoding tag, and it was added in quite a bit after submission, IIRC.)
There's 6 stories in the front page with the vibecoding tag. I think for some periods it's been common to see higher numbers.
There's been 25 vibecoding stories posted in the last 3 days. Very rough numbers, but 8/day? 240 in a month?
It's unfortunate that there's no way to get statistics about "tag usage in the last x period", but if you look at https://lobste.rs/filters, the vibecoding tag has 2000 submitted stories; in the 2k-3k range you have distributed, math, c++, haskell, ruby, unix, browser, games, privacy, compilers... which have been around forever.
You don't need so many people enthusiastic about a topic to flood the front page.
Flags actually have the same issue right now. Their effect on the story standing is very low. As a more recent example, there was that (now removed) story with a vibecoded hardware TOTP thing. People tagged it at spam, and it had +5, -5, but it was still IIRC third on the front page (fourth if you didn't have vibecoding filtered out).
I mentioned this to pushcx a few times, they said the might remove the requirement for the story to also be hidden to affect the score.
Dang, I skimmed through that code and I thought hiding didn't affect ranking. Nice! But that's only combined with flagging? :(
Yup. I think that's really unintuitive.
Comments removing the effect of flags is also unfortunate. I might want to leave a comment explaining why I tagged something as slop, spam, etc. It's something I was doing in the past. Why should that make the flag count less?
When you flag a comment, you're saying this comment is abusive and should be removed. There's no point in interacting with a comment that should be removed.
A better way to explain why you flagged might be to message the commenter.
I'm talking about flagging stories, not comments. I also disagree - there might be context other don't have that explains why I flagged a story.
Here's a good example. The article had hallmarks of obvious slop, but that in of itself doesn't earn you a flag just yet. Upon investigation, the author also had an Generating Content with ChatGPT article elsewhere. It turns out that website was part of a referral SEO scheme!
This firmly moved it into the "spam" category. Even if we did have a "slop" flag, I think this extra context (for why it was flagged as "spam") is valuable.
(I'd also guess that early comment is part of the reasons why other flagged that story as spam, rather than them independently figuring out that it's SEO spam.
I'm talking about flagging stories, not comments. I also disagree - there might be context other don't have that explains why I flagged a story.
I think you can still comment on stories that you flag. It's only on comments that flags remove the ability to reply (and vice versa), if I'm understanding correctly.
My understanding is that the only way to affect a submission's standing negatively is to flag, hide, and not comment.
You are confusing articles written about ai coding/llms/etc with an articles written by AI.
These are different things and not what this proposal is about.
I think they are the same, but I'll come back to that later.
You can consider what I say an example. It doesn't matter what; a few voting/commenting noise has big effects. If a few people vote/comment on "bad" content, even if accidentally, the front page suffers.
I am hugely biased and my perception is likely looking at the past with rose-tinted glasses, but I feel Lobsters has taken a nosedive in quality. (It's still heads and shoulders above anything else, though.) Even if I filter out the vibecoding tag, it might be that my honeymoon with the site is over, but I feel there's less interesting content and discussion than some time ago.
And I blame LLMs. It's all the same. Booster posts, anti posts, stories with endless discussions. They suck the air and energy from everyone. Actually, I even have my doubts on whether anti-LLM content is worse than pro-LLM content. I wouldn't know. You can filter most of that out, but I feel the rest still suffers from the LLM content. (And I feel that because so many people block the vibecoding tag, the problems are just less visible, harder to identify, and there are fewer incentives to address them. But they are still there and affecting everything else.)
Coming back, I was replying to:
There's the question of: if it's low quality, why are people posting and upvoting these submissions in the first place?
I just put forth an example about how anything, including LLM-generated content (not just vibecoding content) can drown the front page with stories that people post and upvote.
Personally, I think the only way to address the problem staying true to how Lobsters has operated until now is to declare LLM-related and LLM-generated content off-topic. Lobsters does not have downvotes and I don't think it's supposed to have sacrificial tags. You could readd downvotes and it might work, but it might just be a dumb, wasteful popularity battle. You can continue using sacrificial tags, but not even making them opt-in seems to be a good strategy.
I've been hesitant to suggest a total ban on LLM-related discussions (covering both the tech, the application, and the social/political outcomes) but I'm maybe kinda slowly coming around to it being not the worst idea in the world.
It would really mean this site becoming an outlier in tech discussions, which would take some time to take effect.
Then there's the question about where to draw the line. Is a submission about a new open source project verboten because it is ok with LLM submissions?[1] How about if something that was previously discussed now uses LLM tech? Is it no longer in scope (for a hypothetical example, the Vim editor has its own tag, what if it went LLM?)
Right now it feels like this stuff is too deeply intertwined in current tech for it to be possible to ban content around it.
--
[1] to be honest most "new FLOSS projects" lately have been low-effort slop, so maybe it's not a huge loss.
I have similar feelings.
The "ironic" thing is that in my opinion, LLMs could be done ethically and usefully. I think the tech is fascinating and magical. But the rich only want to get richer, by all means possible, including lies and deception.
So in my opinion, an entire ban on the topic is the wrong answer. Banning idiots could work- but that would be "banning people I disagree with", which is not a proper way to do things. At this point, I think mandating LLM topics off-topic (I trust the mods to come up with a proper moderation) is the only solution possible. (Anyone can "fork" the website too.) (Allowing people to opt-in explicitly to see LLM-flagged topic might work too.)
at the very least, I personally see this very conversation we're having here as worthwhile. I would not want to ban conversations of this nature, I'm learning a lot about people's perspectives from it.
(the cheesy answer would be to say that discussion of what should be allowed on the site is okay, but I don't think it's the fact that this conversation is about lobsters that makes it worthwhile)
I agree ish but isn't the "low quality" flag supposed to be "do not upvote"?
If we add this are we not effectively adding generic down votes?
I agree that the tag should not be AI specific. There are many reasons why an article can be low quality or inaccurate. The source should not matter. I don't think "low effort" is appropriate because I also don't care how much effort was put forth if it is an interesting article, I feel the flag should focus on the published work, not the process. I think "low quality" or "misinformation" would be most appropriate.
I would still like to ban anything that currently falls under vibecoding currently for a number of reasons, but I’m also not going to get into that now.
I think it’s a discussion we should have soon. I get flagged all the time for even being politely anti-AI, and I frankly worry about being muscled out by people who’ve been consumed by a hype machine.
vibecoding
I’m filtering out that tag since it was created. Out of curiosity, I’ve just disabled it : wow! What a mess! It completely turns Lobste.rs from a nice curated feed into a pool of garbage blog posts/discussion.
Sadly, there are also a couple of moderately-false positives. Like https://lobste.rs/s/w1bsle/lobsters_interview_with_steveklabnik
Yep, I almost missed that one and saw it only because the SQLite migration logged me out of the site... I stopped filtering out the vibecoding tag since.
I’m a little frustrated about how intolerant tech spaces tend to be about genuine criticism of systems. if the people and infrastructure behind a new technology are making life worse for everyone, it feels like we should be allowed to be impolitely against it.
it is impossible to disconnect the system from the idea. without the evil going on behind the scenes the idea would not exist as it does today. the system is still a relevant technical discussion even if it is politically challenging. “ai” is doing untold harm right now and it’s exhausting to see so many people in tech doing free astroturfing work on behalf of companies who see humans as obstacles.
My vague mood of the vibecoding tag is that it's intended for articles that are about using LLMs as part of the software development process. So a carefully-researched human-authored article about using an LLM to (picking a non-contentious example...) translate an old codebase from Fortran to Swift would be a valid article with tags fortran swift vibecoding.
An article that is itself generated by an LLM and is about ... I dunno, the history of JavaScript or whatever ... it wouldn't receive the vibecoding tag, it should be flagged and removed. Because it's not the topic that is undesired, it's the mechanism by which the content was created.
The obvious grey area here is the use of an LLM for machine translation from the author's original language into English. In that case the text is entirely LLM output, but the ideas are human, so it would seem (to me at least) to be OK? But then that opens questions of where the line is, else a spammer could assert without evidence that they used the LLM for translation only, so an enforceable rule might be something like "LLM-translated posts must have a human-authored original".
One other thought occurs to me. LLM-generated output is annoying to read because it mimics low-information text written by human content farms. We're annoyed by LLM output because we assume it has no new information and/or is filled with bullshit. Maybe that should be the flag reason? "This post contains easily disproven claims in the first paragraph and the citations don't exist; flagged as Bullshit" would be both more accurate about the problem and avoid the issue of translation tooling.
Machine translation is older than LLMs and was already useful/useable before.
Text generation also pre-dates LLMs, what of it?
Modern machine translation uses LLMs and does a much better job than the older approaches. DeepL and ChatGPT are both better than Google Translate, so much so that Google is testing an LLM-based implementation (c.f. https://www.lesswrong.com/posts/tAh2keDNEEHMXvLvz/prompt-injection-in-google-translate-reveals-base-model).
AFAIK DeepL has been using „old fashioned“ transformers for a long time. Don‘t know whether/when they transitioned to LLMs and can‘t be bothered to find out since the marketing copy on their site makes my head hurt.
ETA: According to Wikipedia they still are using their own NN implementation for translation https://en.wikipedia.org/wiki/DeepL_Translator
According to their engineering blog, DeepL switched to LLMs some time between 2021 and 2024:
https://www.deepl.com/en/blog/how-does-deepl-work (November 2021)
https://www.deepl.com/en/blog/next-gen-language-model (July 2024)
As someone interested in vibecoding and vibecoded projects, I propose a refinement: If the English text directly linked is AI generated and not clearly marked upfront, it should be flagged. If the code is presented as human but isn't, it should be flagged.
I don't mind a human written post about an AI project, But
I never want to read an AI written post in english, just give me the prompt and save the middleman, my agent can summarize your agent's code better than yours*.
I never want to read source that boils down to a shitty Clos networks of conditionals written by an AI while i'm under the impression that there's a human cared about them. Turns out enormous trees of conditionals are management-requirement-complete, and that's fine, but i don't want to read them under the impression they might have deeper structure.
I also don't like it when a project isn't clear about its code provenance, if it's vibecoded that's fine, just acknowledge it.
It's the same thing as hearing people say they did some renovation that they had a contractor do and I'm curious how. No, you didn't build your kitchen island, you hired someone. If you'd said "I had that done" rather than "I did it" you could avoid an awkward moment when I ask you how the plumbing works in front of that person you were flirting with.
It feels like stolen valor. If Clod did something, credit it upfront. Maybe the flag should be "AI work presented as artisinal"?
*this is a dubious statement. That said, chances are that by the time my agent sees your agent's code, models will improve. From your agent's perspective, mine is from the future. I see no problems with this logic and will not be taking questions :U
Why does it matter how text or code was generated though? The question should be whether there's anything interesting about it or not.
Because LLM cannot add anything of value to the ideas expressed in a prompt.
They add volumes of low density text and a uniform, artificially cheerful and sensationalist voice.
Presenting mostly unedited LLM generated text is a sign of disrespect and should not be tolerated.
Again, if you see low quality content, flag it as spam, I don't see why it matters how it was generated. However, your assertion is false because LLMs can absolutely add things of value by stimulating ideas in the head of the person using them. The exact same way you end up getting new ideas and refining them by talking to another person, or even rubber ducking. The LLM can't intentionally add new ideas, but it certainly can help people develop their ideas.
The sort of a reductionist argument you're making frankly feels like it's in bad faith.
I'm not talking about ideating using LLMs, I'm not sure why you're bringing it up. What you do with your chatbot in private is your private concern.
I'm talking about presenting others with unedited LLM-generated text.
Back to your point, I'd say unedited LLM-generated text carries unique risks: generated claims that are false but convincingly presented. Unedited text carries the correlation that the facts weren't checked either; unedited-LLM generated text is especially low density in terms of information while built to give a feeling of superficial importance, so it represents a specific time-sink.
Lastly, it is a specific form of spam that doesn't look like the rest and can be confused with regular contributions when it isn't
And I've already addressed that point saying that if you see low quality content, then you flag it as spam. Unedited LLM-generated text will almost certainly fall into that category. Hence why I'm not following why a separate category is needed here. We'd reject the type of content you're talking about under existing rules. If somebody posted a link to Weekly World News that was entirely human written, I would hope that would get flagged quickly too.
LLM generated text is "recent". It is novel in the sense that it allows to hide low-effort contributions behind superficially well-written text.
In this context it is interesting to have a separate flag reason, as a means to teach submitters that this particular crop of spam is unwelcome here, and to signal to readers that this is this kind of deceptive content that was flagged.
We're just going in circles here. You're making a contradictory statement when you talk about hiding low effort contributions. Either the contribution is interesting or it's not. Either a person posts content that mostly gets upvoted or they don't. There is absolutely zero reason to have a new flag.
You are completing the circle because you for some reason can't understand that people don't want to see LLM slop even if a bunch of other people think it is "interesting". If you want to be amused by a bunch of LLM slop then, as the commenter you're replying to said, go type in a prompt somewhere in private. Quantity has a quality all of its own; the slop drowns out the interesting human-written posts because there is so much of it.
Because it’s costly for readers to sort out the crap. Because it is a sign of disrespect to have someone read something one didn’t put the time into writing. And because it is utter crap really.
But, readers have to sort out crap to flag the post in the first place. I still don't follow the logic here to be honest. The current process is that if you spot an article that's crap, you flag it as spam. What exactly does a new flag add here?
Meanwhile, the notion that just because somebody used an LLM that they didn't put any time into writing it is itself utter crap really. Some people are just not good at writing, it doesn't mean their ideas have no value. If somebody has a full job as a software developer, and they want to write blogs about interesting things they do at work, I see zero problem with them using LLMs to help format and organize their writing.
Again, what it comes down to is the content of their writing not the tool they used. It's incredible to me that people can't accept that.
Provenance is important. If I see a photorealistic drawing at an art gallery, i'm going to approach it differently than a photograph. If someone says they built something i'm going to approach the conversation differently than if they hired out the work. If i see "low quality" human content, i may offer feedback, AIs can't learn or improve through individual interactions the way humans can. I think I was pretty clear about this.
I never want to read an AI written post in english, just give me the prompt and save the middleman
I actually genuinely love this idea. I honestly would not complain about an LLM-generated article if it started by directly quoting the prompt used to generate it. Then I would be able to restore the balance of effort by only reading the prompt and ignoring the article!
Naysayers will point out that the article's content is thus rendered superfluous, so why bother including it? Okay, let us generously allow the poster to mark their edits from the original LLM output with change bars. If there's enough evidence that the author actually put effort into their own article, I could be convinced to read those parts.
I vote against this. What I care about is the content, not how it was formatted or generated. If there is an interesting piece of code, some factual or thought provoking information, and so on. I don't see why it should be flagged merely because LLMs were involved.
What constitutes slop to me is low effort, superficial content that has no real meaning or thought behind it. Plenty of such content is produced by humans, and this has been a problem long before LLMs have been around. We already have tools to deal with that, so I cannot fathom for the life of me why we'd need a separate category for such content just because LLM was used.
If you see something that's spam, then just say it's spam. It's utterly irrelevant how it was produced. And the converse is true as well.
I don't see why it should be flagged merely because LLMs were involved.
I don't see why content generated at the expense of others should be allowed anywhere where human output is valued, at all. Why do you feel it's okay to for someone to participate in DDoSing small websites (including Lobsters), self-hosted infra by individuals, and ignore copyright when generating anything at all, regardless of its quality, and then claim it's okay to post that content here?
I really wish poisoning LLM scrapers and serving malware to them would become the norm rather than using passive defenses like Anubis. There's no other appropriate course of action when people using LLMs think it's okay to step on others and then talk about how results are the only thing that matter.
Plenty of such content is produced by humans, and this has been a problem long before LLMs have been around.
If that were true, (some) FOSS projects wouldn't consider moving away from centralized platforms like GitHub which are now encouraging contributing slop and have lowered the barrier for entry to such an extent that anyone can generate thousands of lines of apparently plausible code and waste the time of maintainers who do this because it makes them happy, probably not because they're being paid for it.
You're just making a straw man here while ignoring my points. If you're genuinely trying to argue there was no slop before LLMs then I really don't know what else to tell you.
I also have no idea why you bring up FOSS projects contributions in the context of lobsters discussions. These things have nothing at all to do with one another. The original point was whether there needs to be a special flag for submissions that might contain LLM generated content. I don't see any problem on lobsters currently with people spamming low quality content, and if they do that then they should be banned regardless of how it's produced.
There was certainly slop before LLMs, and you were also an asshole if you accepted people posting it before LLMs.
My point here is that LLMs and slop are completely orthogonal. What makes something slop is the intent, or lack of thereof, behind it. I would consider any piece of advertisement, that's been artisanally crafted by organic humans, to be far worse slop than something somebody made with a help of an LLM because they genuinely had some idea they wanted to share with other people.
If you're genuinely trying to argue there was no slop before LLMs then I really don't know what else to tell you.
I don't think I claimed that there wasn't any slop before LLMs. I did claim, however, that the barrier and threshold to create and sling slop at maintainers has almost completely vanished solely because of LLMs. Speaking as a devil's advocate, I don't need to know how projects like Zig, LLVM, Nix, Rust, and Ocaml work to raise plausible looking PRs in all of these projects and waste the time of volunteers maintaining those projects and there's plenty of evidence on GitHub about people doing this.
I also have no idea why you bring up FOSS projects contributions in the context of lobsters discussions
Because FOSS is an extremely relevant topic and domain of discussion on Lobsters and LLMs are destroying the FOSS ecosystem. I don't think LLMs belong here at all, either as threads/links or as people who cheer for it.
Again, I have no idea why you keep focusing on contributions to FOSS projects when the discussion is whether we need a new flag on lobsters to differentiate LLM based spam from regular spam. Seems to me like you're just creating a distraction here to avoid engaging with the subject.
when the discussion is whether we need a new flag on lobsters to differentiate LLM based spam from regular spam.
Oh, I guess it wasn't clear from my post earlier. If Lobsters is not banning anything and anyone who cheers for LLM generated content, the next tolerable thing would be to add a flag for LLM posts. Such a flag can be called "LLM generated", "LLM slop", "AI slop", "AI generated" or whatevery conveys the intent better. I personally lean towards using the word LLM rather than AI. Such a flag should be applicable to any post that has LLM generated or assisted content, regardless of the final quality or "substance" of the post.
I don't think adding flags solves the problem though. People have to spend their time reading, recognizing, and then flagging such content and it would be nice if such a content wasn't there to begin with. LLM content, in my opinion, belongs on corporate platforms, not community platforms.
Meanwhile, for me the next tolerable thing would be if people would just stop constantly whinging about LLMs. It's incredibly tiring, and it's become a form of spam all of itself at this point. If you have an obsession with figuring out whether some content was generated with LLM assistance, then do it on your own time. It's not interesting to me, and I'm tired of these constant debates on the subject. This technology exists, and people just have to learn to live with that fact.
the next tolerable thing would be if people would just stop constantly whinging about LLMs
Sure, I'm suffering from exhaustion when talking about this topic as well. The best thing would be if people using LLMs are ostracized and exiled from community platforms at this point so that people who value human content can enjoy it in peace.
This technology exists, and people just have to learn to live with that fact.
I've read this argument before and it always reminds me of a post from @algernon about not blocking AI bots from our websites. I'll keep doing my best to poison LLM scrapers hitting my self-hosted projects.
You could learn to not be bothered by it. Practise what you preach! Ignore the debates. Disengage from this one, too.
We already have tools to deal with that, so I cannot fathom for the life of me why we'd need a separate category for such content just because LLM was used.
I think OP addresses this:
However, flagging as spam seems to cause confusion. It seems to me that the confusion stems from people who haven't learned to recognize LLM text, and express a confusion that boils down to "this doesn't look like commercial/ad type spam, why the flags?" Having a separate flag reason that conveys "hey this is AI slop" would address this confusion.
If there is an abundance of content that makes it seem like there's substance (which LLMs are good at) then this new flag makes moderation easy. When someone looks a bit closer and realizes it is low effort slop masquerading as substance, no song and dance between moderator and flagger is needed to determine why a seemingly large amount of technical content isn't actually substantive.
And I address what the OP is saying. My point is that I don't think it's worth spending the time perseverating over how content was generated. The only thing that actually matters is whether content is interesting or not. I see OP complaint as being fundamentally superficial in nature. If somebody wants to spend their time figuring out how a particular piece of text was generated, power to them I guess.
Again, I'm going to point out the obvious fact that whether content has substance or not should be evident from the content itself.
What constitutes slop to me is low effort, superficial content that has no real meaning or thought behind it
This problem is undoubtedly amplified by LLMs, I have yet to read an article that is both fully generated by an LLM and is interesting to read/not low effort slop.
Whether or not bad content existed before LLMs is irrelevant, it's a new phenomenon
Previously, I suggested a tag instead of a flag:
https://lobste.rs/s/po97lh/new_tag_suggestion_genai_assisted
I no longer support this suggestion, instead flagging obvious LLM-generated submissions as spam.
https://lobste.rs/s/po97lh/new_tag_suggestion_genai_assisted#c_0vfxye
So, I am quite unhappy about the current situation.
I have been suffering from poor sleep quality for some time now and despite knowing better I have never really looked into that until recently. Being an electronics and programming nerd, I decided to investigate so I threw some SMD dual LEDs to my LCSC shopping cart and it spiraled quickly. Did I use LLMs to build a sleep viewer app in Rust? Absolutely! I am a Rust beginner and it sure helped to get a basic egui app running, to generate IIR biquad filters, to discuss physiology...
If I decided to write about sleep monitoring, would you read it?
But right now, I am myself filtering the vibecoding tag. I have tried to unblock it and skimmed the recent posts. Except for Steve Klabnik interview I think I am not missing much. The block stays.
So, effectively, as things are right now, my own article wouldn't reach me here.
I like to think about LLMs as a group of librarians. They've read a ton about numerous topics, can point out right keywords, but they are not experts on their own. They tend to stay within the domain as human specialists do. Ask current LLMs to code after reading a biology paper and you get what a biologist would write. If you prompt them to rewrite it as a senior developer would, they can do it. But not by themselves. No cross-domain connections.
I am no AI hater, but I do want to read actual people. I don't mind them offloading the gathering phase to robotic librarians. I don't care about snippets generated by LLMs as opposed to copying them from Stack Overflow or transcribed from Wikipedia pseudo code. That's all trivial to me.
I simply want to read a fresh take, with fresh connections, because maybe this time it clicks for me and I learn something.
Also, it seems to me that vibecoding is being filtered by the more picky people in general, so the low-quality submissions make it through. In other words, the community moderation is slowly splintering. It's the most filtered non-meta tag right now, after all.
I would:
Add "Generated" flag for submissions where the human failed to come through.
Rename vibecoding (which sounds quite pejorative) to say ai-tools and treat it with respect.
Now, some people here like to treat any AI use as radioactive contamination. They would likely prefer to get zero AI-adjacent works in their feed. That need is not going to be easy to accommodate in the upcoming years. I think we need to discuss that explicitly.
If I decided to write about sleep monitoring, would you read it?
Yeah. If you actually go write about sleep monitoring, rather than just about how you vibecoded something, I think that would still be interesting.
That's assuming you actually thought about the problem domain, rather than offloading it to the LLM, otherwise the article, at the very least, wouldn't really result in good discussion. I trust that you did, but in general this is a risk with AI-generated stuff, and this is something many people are wary about.
There was some good discussion of this under the recent ajail story.
I like to think about LLMs as a group of librarians.
But to put the full-on AI hater hat back on - considering how my main concern with AI is how they erase authorship, I really don't like that comparison. A librarian is going to provide me with a book. I know who the author of that book is.
An LLM is going to give me an uncredited excerpt.
A librarian is going to provide me with a book. I know who the author of that book is. An LLM is going to give me an uncredited excerpt.
I know. But realistically, even the subject librarian will have a hard time remembering the provenance of the knowledge in their head. They will give you a book they think is most relevant to the subject matter. Not necessarily the book that caused the knowledge to exist in their head.
The models could actually be more reliable with attributions, had they been trained for this. But I think that our current legal system (fuck DMCA) actually makes it safer to train the models in a way that muddies the origins of the embedded knowledge.
In these discussions I always try to consider how I would react to a fully libre, consentual model. That helps me figure out what I am mad at specifically, rather then conflating the issues of the technology with workings of capitalism or geopolitics.
Now, onto the harder part.
I think that would still be interesting.
This is a hell of a way to discourage people from sharing. I sincerely hope that was not the intention.
That's assuming you actually thought about the problem domain, rather than offloading it to the LLM ... and this is something many people are wary about.
Get off your high horse. Just because some young-uns felt excited from accomplishing something (with the help of AI) doesn't mean you get mean and manipulative. It means you acknowledge their efforts and gently steer them towards growing again.
In these discussions I always try to consider how I would react to a fully libre, consentual model.
I am not aware of a single such LLM. Do you know any?
(For what its worth, I also wouldn't lean too hard on el problema here. Even if we lived under fully automated luxury gay space communism, plagiarism would still be wrong.)
I think that would still be interesting.
This is a hell of a way to discourage people from sharing. I sincerely hope that was not the intention.
...I don't understand what's the issue there? "That would still be interesting" - as in, compared to something that did not involve AI at all.
To try and clarify the second part - my point was "people might use the presence of any AI assistance as a signal that the article might be low quality". The story I linked has a comment directly expressing that sentiment with 9 upvotes as of right now. That's what I was referring to. That's why I linked that thread.
Apertus claims to be (discussion, where apparently that is rebated).
I tried it briefly, was exceedingly easy to make it hallucinate.
edit: to add... I think if a true non-profit said "we're going to make an LLM to make alt text and other accessibility things, and we will cap profits to x%, make everything transparent and OSS, etc." many people (myself included) would happily contribute to the training set.
The linked comment mentions Comma, so I just want to mention to mention that it still blatantly violates FOSS licenses, so I'm skeptical of the motivations behind it.
I think there's a world of difference between grabbing everything "accessible" online, promising to respect "opt-out" requests eventually, and actually getting affirmative permission to use material from its creators. Letting people opt-out is better than nothing, but in my opinion not nearly good enough.
If you did require explicit consent, I'm sure some people would be willing to agree to contribute, but the unwritten implication is always "we don't think we'd get enough training data if we asked for it, so we don't ask."
"Yeah. If you actually go write about sleep monitoring, rather than just about how you vibecoded something, I think that would still be interesting."
This is the whole problem of the vibecoding tag. It is a generalization.
I use agentic coding tools too and I can guarantee you that I meticulously design and review generated code. It is far from "vibecoding" where i say "build this and that" and never read the code. And yet according to the rules here it is all vibecoding.
Reality is, we can now individually choose that this way of working and these tools is now part of our craft and denigrating every use of agentic coding tools to vibecoding does not show much respect to your "colleagues" here. It has a lot of "real coders do not use IDEs" vibes.
I fully realize that "AI" in general is the most divisive topic in our industry at the moment, but I think mutual respect, having an open mind and not putting negative labels on things is the basis of a healthy and happy community. At least we could try.
dude, these tools are known to directly plagiarize the works of others, without any attribution. this is something i think is unethical, and you have no way of avoiding this when using them on a large scale. why wouldn't i put a negative label on that?
Thanks for the defense. I am not a big agentic coding believer, though.
But yeah, it would be nice if you could concentrate on the outcomes and talk about what specifically we want to see instead of labeling everything vibecoding and dismissing it as toxic.
I vote against it, unless it is called "low effort/quality" or something. I don't care if it's LLM genrated or not. I care if it wastes my time.
I've seen people become automatically suspicious about posts that use mdashes and bullet points.
That saddens me, because these used to be indicators of high-effort posts. I've used to deliberately insert non-ASCII punctuation, and structure text for the "F" scanning and inverted pyramid. Now counter-signalling is needed to be taken seriously online.
And what is the recourse here? What if I am not great at writing english and I use an LLM more like a spell/grammer checker but 99% of my work is original. Or what if my writing style just looks like "AI generated". How are real people going to get an opportunity to get their voice heard. How do you convince anonymous downvoters that you are a real human. Where do you go? What is the process?
What if I make a genuine effort to write an article and the group here decided it is not real. Put yourself in the shoes some somone in that situation here and think about how they would feel.
This is a textbook example of what tags are for and flags are not for: "Lots of people are submitting and upvoting a category of content that others don't like."
We already have an off-topic flag. If a post is on-topic, not spam, etc., and getting upvoted a lot, then flagging it makes no sense. The upvotes indicate people want to see it. If others don't want to see it, that's exactly why we have tag + filter: so it's only seen by those who are interested. (If you want all LLM-related posts to be considered off-topic for Lobsters, that's a different proposal.)
We don't need a new flag, we need to replace the vibecoding tag. Neither people who want to see LLM-related content nor people who don't want to see LLM-related content seem to be happy with how that tag is working out.
I've been sitting here thinking about this for a while and I'm not sure that lobste.rs has any good path forward other than rolling this LLM slop into the existing mechanisms for "low effort" submissions, as a couple of other commenters have discussed. If your goal is to take links from anywhere on the web (as it is for this site) then you get what you get. And worse, any mediocre submission is going to have half the comments arguing about whether it's slop or not. This is a relative weak spot at the moment since there is no flag for it, so I anticipate stronger mechanisms will eventually be needed: either a downvote mechanism, or a penalty for users who submit weak submissions in general.
I feel the more robust solution is on the author side. I'm imagining maybe a sibling site to lobste.rs which has the same invitation system but only self-authored posts are allowed, and a condition of inviting somebody is that they will only ever write in their own voice, share posts in their own voice, and clearly mark anything that's AI generated. Or equivalently with a different tech stack, a mailing list where only PGP-signed messages are permitted which validate in a web of trust based on the same rules: people vouch for each other's commitment to writing as humans. With the ability to run LLMs locally, nothing is watertight, but something like this would be a much stronger arrangement if you object to reading LLM output.
Alternatively, people could stop perseverating over whether something is LLM generated or not, and focus on whether it is in any way interesting. The problem with threads devolving into endless discussions of slop is with people making the comments as opposed to the submissions themselves. People on here could just behave better, and stop making vapid comments.
There are two schools of thought here. One is that interesting content is interesting content and submissions should be evaluated on that basis. The other is that human-typed output is inherently superior, for various reasons - originality, authenticity, or simply that it has more meaning because somebody took the time to type it. I assume that people who browse lobste.rs today are a mixture of both groups.
I disagree with your suggestion mostly because the economics have changed. I have access to several chatbots right now who can give me 500 pretty interesting words about whatever tech topic I choose, in a matter of seconds. So do most people. Everybody can do this privately, right now, for free. Nobody particularly benefits if I copy-paste one of these responses into an HTML page and share it on a link aggregator. It's just a crass form of curation.
It is the human-written posts that have unique and creative insights, that describe the human experience of building/using technology, and that provide the raw material for the LLMs. Those are the posts I care about above all, for quite rational reasons, even if it means rejecting "more interesting" posts that were produced more algorithmically.
If human produced content is inherently superior, then there wouldn't be a discussion in the first place because it would be obvious which is which based on the quality of the content. The mere fact that this discussion is happening indicates that human content isn't inherently superior or obvious. I see zero evidence to support the notion that organic homegrown human generated content has some special quality to it.
And yes, anybody can use chatbots to generate content. But lobsters is a moderated community, and if somebody starts spamming then they can be banned on that basis.
Let me be clearer: in my view, human produced content is superior (something I want to read) even if it is inferior quality (technical insights, quality of prose) because it has elements that are fundamentally unreplicable when an LLM writes about the same topic (lived experience, dreams, the mannerisms and foibles of a real person).
It is not my mission to make you agree with me but it's clear that if we have such different positions on the value of posts then we're not going to easily find common ground about how link aggregators should rank and moderate submissions. That's a useful finding in itself. Which side is lobste.rs on?
Elaborating on this: when a person writes, I may not know what that person’s exact intent is, but I know that it was produced with intent of some kind, that someone is attempting to be understood in some way. LLM-produced text has the surface appearance of intentional writing without any of the core. The prompter is ceding responsibility for word choice, structure, personal voice, and so on. The idea that I should treat that kind of text with the same scrutiny I give to something written by a person, and to expect me to ignore the context under which a piece of text is produced? I find that deeply insulting, anti-intellectual, and completely out of step with the kind of environment lobste.rs has often been and could continue to be.
Sure, we have a difference of opinion on what constitutes interesting content. I guess we'll see where majority opinion lands here.
Thanks for putting this into words! I find the philosophy behind it quite interesting: text as a standalone artifact that should be judged in isolation vs. text as something social in which the human beings around it are as important as the artifact itself. I'm definitely in the second camp.
Well, no, I don't think it's actually obvious from the quality of the content whether I could have gotten that content by typing a reasonable prompt into an LLM instead of the linked page. That difference matters for two reasons:
Posting content that could be gotten from an LLM to lobste.rs is duplicative, since anyone who wanted that content could have gone to the LLM instead. There's very little to learn without seeing the prompt that generated the content so that we can map the initial idea onto the generated artifact.
Very often, the reasons the content is inferior are not present in the content itself. It's often present in the relationship between that content and a context surrounding the content, such as reality. Code may function in a demo but fail in common real-world circumstances; an essay may reference papers that were never written; and so on and so forth. Figuring out when these are problems in some piece of content requires expertise, which readers looking to learn about unfamiliar topics definitionally do not have. Content generated without grounding in reality is poison to those readers.
This is true even if the mistakes aren't part of the core concept of the post. This may actually be the worst situation since experts are likelier to focus on the main point, which could be genuinely useful for someone who knows enough about the topic to discard the cruft. When dealing with human authors in an informal knowledge-sharing context, it's typically considered disrespectful to do the kind of public nitpicking that would be necessary to help prevent misunderstandings from creeping in based on peripheral nonsense. Knowing it didn't come from a human can help signal that those norms don't apply.
AI generated: text or code that was substantially not authored by a human mind trying to communicate an idea, but rather by having an LLM expand a prompt into a larger artifact
I think that LLM-expanded text and LLM-written code are two very different things, and that the former is objectionable in a way that the latter simply isn’t. Written text serves two purposes: the obvious one is to communicate an idea to others; the less-obvious one is for the writer to grapple with the subject. An LLM-expanded text contains no more original thought than was in the prompt, but is longer, so it wastes readers’ time, but it also doesn’t require the writer to grapple with his own thoughts on the subject. It’s an attention DOS.
OTOH, code is completely different: the point of code is to communicate with a chip in order to achieve some sort of effects. This is obvious with a binary executable, and LLM-generated code is no less useful than a binary-only piece of software: if it achieves the desired effect (big ‘if,’ sometimes!) then great. I doubt very much that any one of us has read a significant portion of Firefox, Chrome, Linux, ls, vi, Emacs, GNOME, KDE, GIMP, Inkscape and (not ‘or’: ‘and’) X11, but those programs are still valuable to us all. Likewise, an LLM-generated program in a high-level language can be a valuable artifact.
I think that it’s also true that the use of LLMs to generate running software involves a lot more human judgement and interaction that to generate text, and that to just dismiss that human effort is … dismissive.
Ps.: I have few problems with LLM-edited text — it doesn’t strike me as particularly different from using a spell- or grammar-checker. And I completely hate it when folks think em-dashes are some guaranteed indicator of LLM-generated text. I love my em-dashes!
Pps.: Perhaps the thing I like least about lobsters is the dismissive way in which some perfectly legitimate content is disappeared. I often see stories which get removed in my RSS feed: in my opinion a lot (perhaps not a majority, or even a third, but still a lot) of them are perfectly good, and I’m glad that I had the opportunity to read them. I don’t think adding another thought-terminating flag is a good idea.
OTOH, code is completely different: the point of code is to communicate with a chip in order to achieve some sort of effects
I fundamentally disagree. Code is to precisely communicate the configuration of the chips to humans, and the act of writing it requires the writer to grapple with their thoughts, distilling them into a clear and understandable algorithm.
Indeed. Code should be written by human because it is for human. Else, why not generate binary code directly?
Also, the person assumes that because he didn’t read source code for Firefox, Linux, ls, vi, Emacs… then nobody does.
Well, it’s false. Lot of people do it. That’s even the purpose of Open Source. And it often happens without premeditation : you need something and, by reading the source code, you are able to solve your problem.
I’ve read the source code for "less" and "file" in the latest years. Even contributed to "file" by identifying a bug. It was not planned. It just needed to be done. Open Source AND readable source code allowed me to do it quite easily.
Vibecoders are just inexperienced coders who think they have found a magical shortcut while they still haven’t managed to understand what "code" means.
Code is to precisely communicate the configuration of the chips to humans
My very next words were: 'this is obvious with a binary executable.’ That’s code, too, and it communicates nothing to a human being on its own, and even with a disassembler it is typically a pretty painful read. It’s not source code, but it’s still code.
Re: your PPS
Any story that's removed should have an entry in the modlog. I don't find much there that's unreasonably deleted. If you have issues you should raise it with the mod team.
I'm strongly in favor of this. I'd also support the slightly more general slop or low-effort as suggested in another comment thread.
I’m conflicted about this. On the one hand, I’ve also seen a ton of slop submitted and I find it frustrating. I go back and forth about muting the vibecoding tag, but there’s just enough “real” stuff there that I haven’t yet.
One of my frustrations has been seeing articles relating to actual machine learning work — which I’d expect to see under the “ai” tag —tagged “vibecoding” instead. Even when it doesn’t relate directly to LLMs at all!
So if blocking AI slop completely led to a better signal-to-noise ratio for real technical contributions, I’d love it. I worry though that those same ML articles would end up blocked in the same fashion, rather than floating to the top.
"IMO, the feed is being DoSed by LLM-authored text and vibecoded software"
Would be nice to see some actual evidence of that. Not saying it is not correct, but it would be good if someone actually took say the last 100 articles posted to show if the feed is indeed dominated by AI generated articles or not.
I'm asking this because the number of times I read something here and my spider senses go off "this is not written by a human" has been extremely low. That is my personal experience.
This is not Medium where that is extremely prevalant. I have a feeling that Lobster members, the first filter here, are actually quite good at not posting low quality articles here.
"Anecdotally, the community's immune system is currently rejecting these in two ways: adding the vibecoding tag to anything that's been touched at all by an LLM"
This is not correct that tag has been liberally applied to all kinds of articles from people writing about research, software engineering practices, practical example of using LLMs and even just a tiny mention of using LLMs for coding in a project. That tag has completely lost its meaning and using it now also for content that has been generated with LLMs will add even more ambiguity.
Can you show proof that this text wasn't generated by LLM?
That said, perhaps some creations are obviously AI-created and that was the rationale for this flag. If AI-generated text is indistinguishable from something created by a human, nobody's gonna think of flagging it, I guess?
In general, I agree with this proposal, though I worry that the line between "what's substantially human authored" and "what is vibecoded" is going to be pretty fuzzy. Having been part of a recent case study, it seems like there's a contingent of folks who think any-LLM-contributions-means-vibecoding, so how would we define "substantially human authored" in terms of tagging?
One thing I'm concerned about is that it seems like the hygiene of even labeling LLM contributions is going to be a mixed bag. Here Russ Cox argues that LLM contributed patches to the Go project shouldn't be tagged as Co-Authored by an LLM (with a strong argument)
I think how about a limit on post length? I see these 10,000 word monsters and just want to runaway screaming. I know what people need to say things, but if something is like 10,000 words it means they don't know how to write and so it's either AI or crap. Also, I think there should be a nice video tag too. And a PDF one because it kind of screws with my my reader! Perhaps a little thing that goes out in previews, the number of words in a particular post and if it's a 10,000 word monster require some kind of annoying process.
