Post

it is stunningly easy to manipulate LLM chatbots into saying whatever you want - i noticed this the other day when LLMs I was experimenting with pretty instantly started regurgitating my own Bluesky post, almost as soon as I'd made it, identifying the sourcing of an image.

bbc.comI hacked ChatGPT and Google's AI – and it only took 20 minutesI found a way to make AI tell you lies – and I'm not the only one.

19/2/26, 6:58 pm19 reposts105 likes

Mae B.

@maebee76.bsky.social

19 Feb

Dude this is literally what my tween kids do when they are bored. They compete with their peers to manipulate chatbots into saying bizarre, wrong, and horrible things.

Nobodyatall

@criquetgen.bsky.social

19 Feb

mjfgates

@mjfgates.bsky.social

19 Feb

I *guess* she still counts as related, anyway.

Parkscope Joe

@parkscopejoe.bsky.social

19 Feb

How do LLMs handle new information when they’re not retrained? Is it attempting to pull and read from the web?

Faine Greenwood

@faineg.com

19 Feb

Yup en.wikipedia.org/wiki/Retriev...

en.wikipedia.orgRetrieval-augmented generation - Wikipedia

Parkscope Joe

@parkscopejoe.bsky.social

19 Feb

Ooooooooo some fun reading

Virginia Petit

@virginiap.bsky.social

19 Feb

which is why the LLM's are posting disclaimers and people should question everything AI slop feeds them. Meanwhile Elon suggests we should feed medical information to Grok 🙄