Hacker News Reader

Comment by anonymous908213 | Hacker News Reader

Bot account. Last active with two posts in 2017, then started spamming today with the usual HN spambot formula. As usual, doing the weird thing that only the HN spambots do where it's almost always exactly three paragraphs of 1-2 sentences each. Not sure why they prompt for that formula but it makes it very easy to spot out. For me, I guess. This is voted to the top and the bot has over 50 karma so apparently most people are unable to detect LLM spam even when they make it as obvious as they possibly can.

View on HN

As a master in commenting I would argue that this is quite smart. Indeed this might be a bot.

Or maybe it's not, who knows? It's sometimes hard to tell with comments.

It's just the times we live in, uncertainty is a given, most of the time we don't know. I guess we'll have to make do.

Gotta try harder than that. There's a certain length to their paragraphs as well, you can see from their post history the general shape of their messages that will cue you off that you're probably about to read spam before you've even read the first word. Also abundant LLMisms. Four "not X but Y" and three lists of three in only five sentences. The sheer density of LLMisms is absurd. Not to mention that vague sense of incoherency despite overall being a grammatically correct series of words. There are so many tells it's criminal people aren't catching it.

I wish I had a super-upvote.

"OP here" or "Author here" in posts is an obvious one when you see it. (In both cases, it's completely redundant!)

As for comments, regurgitating the input too closely and overusing abbreviations are highly indicative, yet many commenters don't notice.

See for instance (slop warning) https://news.ycombinator.com/item?id=46986273 and https://news.ycombinator.com/item?id=46965103.

I would suggest sending this as an email to the mods instead of a reply, what you're doing is practically a bug report to the bot owner right now. The resulting discussion isn't super interesting either (IMO).

I am not signing up to a mini-moderator who has to e-mail dang 5 times a day. I want to not see these messages in the first place, and for that to be possible, the community needs to learn and be able to recognise this for themselves so this spam can be flagged and killed on sight.

If the bot owner happens to waste their time reading the responses to the dozens of comments one of their many spambots made, and improves the bots as a result, so be it. They're already winning the war as it stands, not like things can get much worse. I'd like to at least try to make an effort to make things better.

Maybe the actual answer is that I just need to stop using HN, though, since the spambots are taking over the site and yet people are more concerned with the people pointing that out than the actual problem.

OK, I sent the email myself. You don't have to do it. Just once, not 5 times per day. Agree about community awareness, but I think emailing the mods is more effective than responding to one of the 1000 comments this account is making.

Thanks for doing that because I had no idea this was happening. I've banned the account now and flagkilled the 30+ comments they posted today.

you missed one: https://news.ycombinator.com/item?id=46998241 ;-P

two actually: https://news.ycombinator.com/item?id=46988519

Those were killed. What made you think they weren't?

i most certainly saw them as neither dead or flagged.

now i see three other comments as alive:

https://news.ycombinator.com/item?id=46997839 https://news.ycombinator.com/item?id=46996890 https://news.ycombinator.com/item?id=46992786

i know they were flagged and dead before.

i can make a screenshot.

could those comments be reactivated because people vouch for them? or is there some kind of bug?

That's right—some users vouched for those 3 after we'd killed them.

That makes sense, because if you look at individual posts in isolation you might think they were unfairly killed. One needs to look at the posting history of the account as a whole to understand what the issue was.

(Btw, thanks for explaining the context in those other threads.)

I've had repeated comment and email exchanges with you over the years over whether or not an explicit "killed" or "dead" indicator on accounts and their posts/comments should exist. I understand the reasons against this, and arguably they'd be more relevant in the case of detected bot accounts (the indicator would be yet more training data, assuming feedback training).

I've also been experimenting with my own indicators for specific accounts based on my own criteria and interactions which I've found useful (applied through my own HN tweaks). E.g., if I see an explicit mod note that an account has been banned, I can mark it as such myself, sparing confusion.

How HN can implement a Voight-Kampff test becomes an increasingly relevant question.... One of several HN needs to consider with increasing urgency.

(Three 'graph pattern ... is again noted...)

more important than an indicator it would probably be useful to somehow disable vouching for a killed account. i don't know if it is possible to set the flag counter to some high number so that vouching simply has no effect or to an invalid number like -1

That's antithetical to how HN has operated in the past. Vouching for deads is fair when the account is an actual human, and happens to post valid content. I do this occasionally myself (I read with "showdead" on), though not especially often.

See, e.g., <https://news.ycombinator.com/item?id=31525284> also the FAQ: <https://news.ycombinator.com/newsfaq.html#cvouch>.

Accounts killed for spamming AI content seem to me to violate that premise, and an unvouchable kill does seem appropriate, especially where it's not immediately evident to the casual reader that an account was killed for posting AI content.

I'm thinking of how I'd like to indicate such accounts myself, and am leaning toward adding a robot emogi via an ":after" CSS rule.

i think we are actually agreeing. i am not talking about making kills unvouchable in general but i am suggesting how an unvouchable kill could be implemented without to much effort. the unvouchable kill should of course be only applied to appropriate cases, it's not meant to replace the regular kill.

Yes, we are in agreement here. I was simply noting what HN's past policy and rationale have been.

LLMs change the calculus somewhat in making automated bot-posting far more viable. It's clearly already a problem. I suspect that moderation policies will have to adapt to this. There's also the fact that such a change would make AI-banned discernable from normal bans, in that AI-banned accounts would not have vouchable comments. If explicitly noting AI banning isn't adopted on the basis that this would provide information to either the AI or its operator of the fact / nature of the banning, the absence of a vouch option would reveal the fact regardless.

(A relatively small example of changes we'll see induced by LLMs in the larger world as well. Interesting times....)

my suggestion would not remove the vouch option, it would just make it ineffective. people using it would not notice. the system would still indicate that you vouched for a message.

Fair enough. Email the mods! ;-)

Thanks - this was a good catch, and it makes steam come out my ears to see any account abusing HN like that.

You don't have to email us, of course! But please understand that we're on the same side. We don't want to see HN overrun by generated comments (a form of spam) any more than you, or other users do. Remember that tomhow and I were avid users of HN for many years before we became mods.

All: generated comments and bots aren't allowed here. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que.... If you notice an account that appears to be consistently pattern-matching to this, and have a minute to let us know, we'd appreciate a heads-up at hn@ycombinator.com. We don't come close to seeing everything that gets posted here, but we do monitor the inbox closely (including fishing through the spam bin for real users) and we take these reports seriously.

As someone very engaged in reading HackerNews comments I observed that your comment and the comment you are replying to have that exact same format of three short paragraphs.

The people behind these bots most certainly found that many engaging, authentic comments follow this clever pattern. It is also worth noting that such comments are remarkably digestible – due to their brevity and decomposition into even smaller logical and lexical parts – and swiftly read, requiring only a very short attention span and little intellectual investment from the reader.

This makes me very curious about the statistics on how HackerNews comments are structured and how well different formats of comments perform in the community. I would be thrilled to dive into the data and might write a neat program to analyze this sharing the results with the community.

The data you'd need for that is all available through both the HN Firebase API (which is a bit antiquated) and Algolia's HN Search API. If you find anything interesting, definitely let us know!

Incidentally, both comments were edited after the fact. Neither my nor Dang's comment were originally of this length but we both made edits that ended up there. I did notice the irony :)

I just wish they would use bigger models. These ones always write stuff that makes no sense. Like it says "the same principle applies to mobile games" then describes a different principle

There's some bots on HN who write much more coherently and get a decent # of upvotes. I was only able to catch one because the comment started with something along the lines of "Here's a smart response for a technical audience about _____"

lmao, got a link?

Sadly it was edited, but this comment used to be clearly responding to a prompt: https://news.ycombinator.com/item?id=46805888

@dang what was the original comment?

I've lost the plot - what comment are you asking about?

(Btw, @dang usually doesn't work - it's better to email hn@ycombinator.com.)

Thanks. I upvoted one of its comments yesterday; i.e., it fooled me.

i saw a flagged comment and thought who the hell would flag that? it makes no sense. was going to vouch for it, but i always check the comment history which is where i noticed that all comments had a suspiciously similar pattern, and i found this thread.

some indicator that an account is banned would be nice for those who have showdead active.

one thing that is disturbing is that the comments of the bot are all lowercase. is that a bot feature now? are they doing that to appear less like a bot?

do i have to change my style to avoid looking like a bot? or is changing my style going to make me look like a bot?

i feel trapped.

No, most of the spambots use proper casing. Whoever is behind this does prompt different bots to generate output in different styles with the intent to appear more conversational/human, so each bot has its own "personality" and typing style. An example of another style here: https://news.ycombinator.com/item?id=46885996

Any idea what the endgoal of such bots is?

Karmafarming and then astroturfing?

Astroturfing seems to be the goal. I notice sometimes there are threads in which numerous of these bots converge to promote some dubious product being shilled.

Do you check everyone's hn account history?

No, I check account history when the text is obviously LLM-generated. I mostly point it out because if I don't make it abundantly clear how obvious the spambot is by its behaviour, I will get people telling me that it could totally be a human and that I'm making a false accusation.

Might be because while it's not high quality, it is ultimately a constructive comment. So this might be a case of https://xkcd.com/810/

I suppose 15 years ago it would not have been possible to predict how fucking irritating these comments could be. I would rather have the old spambots back at this point. Please, spambots, go back to trying to sell me penis enlargement pills, I beg of you.

Now this is a definite human comment, shoehorned in xkcd links are a huge human tell.

Just doing my part

Reply on news.ycombinator.com