logoalt Hacker News

ultratalkyesterday at 2:50 PM3 repliesview on HN

Out of curiosity, has anyone noticed a non-negligible presence of bots in threads on HN? I haven't, but I'm not sure if that's because I'm bad at spotting them or because HN is good at getting rid of them or because HN is a niche platform.


Replies

armchairhackeryesterday at 5:48 PM

Yes, they’re very identifiable. New or resurrected account makes multi-paragraph comments on random topics with “insights” that read like AI, even if they don’t have em-dashes or “it’s not X it’s Y” (and sometimes they do).

Fortunately and in fairness to this site, they’ve become rarer, and most seem to be flagged within hours. Usually I look at the comments to confirm, and most are already dead.

show 1 reply
kartoffelsaftyesterday at 4:28 PM

I made a post here a bit ago where one of the few replies I got was one of these conversational ad-bots, albeit on the more obvious side. It was getting flagged which gives me hope that HN is good at filtering it, but I also mildly worry I'm (or we're) just missing it when it's subtle. I do suspect it's a huge volume in terms of comment count either way though.

jerfyesterday at 3:40 PM

I have suspicions but there's fewer signals on HN available to the general public so it's harder to tell.

Well... to be more precise... I'm abundantly positive there are bots and shills here in a general sense. But when it comes to identifying specific accounts as bots or shills, it gets difficult. Yeah, a lot of us have gotten pretty good at identifying the "default LLM voice", but it is trivial to kick it out of that.

I have done some formal writing with AI, and I always feed it a sample of my own writing to emulate. It doesn't do it perfectly. For instance, I'm a semi-colon kind of guy and it still em-dashes without more explicitly instructions to avoid them. But what comes out the other end would definitely pass most people's "default LLM voice" sniff test; it eliminates most of the tells [1] people look for. (I just checked. The resulting output may actually be "better" at avoiding the tells than my own actual text...)

The upshot of all of that is that we are approaching a point with the current AIs that with just a bit of clever prompting it may take many, many kilobytes of text for someone to form a justified (!) opinion that some set of posts is actually AI.

[1]: https://awnist.com/slop-cop