(a bit more about this at https://news.ycombinator.com/item?id=47056384, with a reply from the OP)
So maybe some sort of filter like that? Only show it to those kinds of accounts at first?
The downside is that if that group isn't big enough you get a lot of groupthink, but if your sample is wide enough, it can be avoided. To be honest, I don't recall why we stopped doing it.
what I’m seeing is new or sleeper accounts that have been idle for over a decade with low (<99) karma getting into comment circles. Over the last couple of weeks i’ll see several top comments on articles with back and forth between other similar accounts… it’s got to the point that I check a user habitually before I even bother reading… and I have never hidden so many comments before getting to something substantive in the comments…
Like many here, I don’t wish to limit new users, but this does seem from my armchair perspective to be a pattern to be on the look out for.
I've noticed this kind of behavior on Reddit but never on HM
Rumor had it that there is also some kind of social-network metric detecting when socially adjacent accounts (or alts) are engaged in astroturfing, the practice where a small cabal tries to pass themselves off as a broader grassroots campaign.
Flip that around though and the same metrics might allow new accounts to be meaningfully vouched for by existing ones.
A few people in these comments seem wildly confident that it is written by an LLM. If anything, I hope it was written by a human as an elaborate troll to trigger these so-called immaculate LLM detectors.
Would seem to require some discernment to classify. Not all assistive use is slop.
> Not all assistive use is slop.
That's right, and the key is to discern which posts/projects are interesting.
It's obviously says LLM to me at first read through.
I suspect that:
a) less people are willing to expend a bit of energy to notice LLM usage given how much of it is. ("we've lost" theory)
b) that people are losing the ability to detect LLM submissions. ("we're cooked" theory)
or c) that people don't care about the use of LLM. ("who cares" theory).
Personally I've been feeling less invested, because it seems as if most users don't care and even the main users of the site don't notice it.
I should clarify and revise my thoughts and initial comment. I do not think that not being able to detect it leads to lack of care. I actually think that many things have passed me by and in the future this will be even more as LLMs improve ("we're cooked").
As to "what do we do when we spot it" - you hit the nail on the head of the feelings I felt as I was writing the comment. What do we actually do, what can we change and should we attempt futile things?
And even the example dang gave - the actual submission as very good. Is any amount of LLM use okay and what's the level? I use LLMs at work but I don't like writing readmes or blog posts with it. But others might like writing code at work by hand and don't like writing text so use LLMs for that. Maybe I lower my expectations!
Who aside from Fred fucking Durst writes like that?
Ugh... Clearly llm generated. This is how internet has become. 90% of posts are variations of tropes like these.
> I am sooo tired of statements like "No x. No y. No z." and then optionally "Just Foo.". Who aside from Fred fucking Durst writes like that?
Well, the simplest automated method would be to run the post and comment together through an LLM with a prompt that's roughly:
"Is this person claiming to be the author or co-creator of the work discussed in this submission?"
Only green accounts subject to it. I predict you'd probably have a very low false positive and false negative rate.
It's of course a terribly slippery slope. My perhaps overly-cynical take is that once the infra is place some of your bosses would be prone to eventually abusing it.
Personally I'm here for it: Dang, moderator turned whistleblower—on the run from dark VC money—in a race against time to save freedom. Still working on a title for the film.
McDang II: You Have the Right to Remain Shadowbanned