Taking over subreddit with a small GPT-3 army turns out to be surprisingly easy

https://old.reddit.com/r/singularity/comments/wa9enf/it_took_me_1_day_to_create_a_program_using_gpt3/

by EGreg • 4 years ago

49 23 4 years ago

old.reddit.com

Scary stuff, and it makes me think that the days of anonymous online forums are coming to a close. Even HN, unfortunately. I'm not sure how you stop this from happening without increasingly invasive KYC-type measures.

AnimalMuppet• 4 years ago

I wonder if it's already somewhat pervasive on HN.

I see posts that don't quite make sense. Well, we have people here whose first language is not English, so you can't expect perfection.

I see posts that have a definite agenda, or spin on reality, or bias. But humans have those things, too.

But I see the volume of this going up in the last couple of years. I wonder...

ch4s3• 4 years ago

There’s a certain topic I see those on, but I think they’re a mix of state actors and rabid partisans wrt the issue in question.

EGreg4• 4 years ago

Similar from 4chan: https://www.youtube.com/watch?v=efPrtcLdcdM

And they only found out 1 bot out of 9!!

Klonoar• 4 years ago

Related, ish: Twitter bots potentially posting with ChatGPT-generated contents https://twitter.com/levelsio/status/1604841600416624642

quantified• 4 years ago

For malicious actors, life is getting easier. We may find social networks or online circles where being properly introduced becomes a thing again.

etrautmann• 4 years ago

Isn’t that already a thing? In my world all social networking is via group chats and texts threads.

karmakaze2• 4 years ago

That may be true on Facebook, but the same when reading Reddit, Twitter, Stackoverflow, etc.

How would you tell how much to trust what you read?

stjohnswarts• 4 years ago

or to have meaningful conversation you'll have to register so identification and then be under scrutiny for trollish behavior, at least to minimize bad actors, you'll never get rid of all of them.

karmakaze2• 4 years ago

Or web-of-trust as done in PGP to vouch for users you didn't directly know.

yucky• 4 years ago

Won't be long until the Government puts themselves in charge of this, making sure you only communicate with approved sources. For the children, of course. And terrorists, or something.

sys_64738• 4 years ago

I read a great comment but forget it completely. It basically was the human algorithms are generally predictable and will be easy to simulate on AI for the long run. The real concern is unpredictable AI that isn't human based. Something like that.

l0b04• 4 years ago

tl;dr, but not surprising at all. Forums (including HN) haven't evolved to deal with motivated saboteurs yet. Basically, any content with correct grammar and spelling have already cleared the most useful moderation barrier, which historically kept out the inarticulate lunatics and primitive bots. Probably the only way to deal with this is networks based on invitation (like lobste.rs) and a twitchy ban hammer.

EGreg4• 4 years ago

Doubt that would work to keep bots out.

Anyone can invite bots or hook a bot up to their own accounts. Then what?

l0b04• 4 years ago

Then anyone who is higher up the invite ladder can ban them, and everyone they invited. Shouldn't take long to weed out the few miscreants.

welshwelsh• 4 years ago

Making a popular public space invitation only or super ban happy means the attackers won. In that case they have successfully made the space inaccessible to the public or made it so that normal users are now getting banned.

l0b04• 4 years ago

Nah, the public can still read everything on there. Just not participate, unless someone already in the community deems them trustworthy.

MrMan• 4 years ago

if I owned a social network I would start trying to differentiate on quality of generated content instead of demonizing automated user agents. I think its definitely a cant beat em join em type scenario

EGreg4• 4 years ago

If some are better then they are all better and dwarf human content

Then when they bring capital too, why have humans at all? What would the humans add to a typical discussion?

jokoon• 4 years ago

I have always been thinking that AI could generate authentic looking comments, to conclude that Reddit would mostly be a gigantic playing field for astroturfers.

Taking over subreddit with a small GPT-3 army turns out to be surprisingly easy

Related Stories