> Australian researchers told an AI model to act drunk and found it was more likely to break rules or divulge information it was supposed to keep private.
https://www.abc.net.au/news/2026-09-29/drunk-ai-testing-unsw-research/107208104
2 Boosts
0 Likes
@davidgerard Honestly the most surprising thing about this story is that they felt the need to include this sentence…
Welp...
Training the bot on all those drunk texts sure worked out well, eh?
🤣😂😁😁🤪🙄
@davidgerard ok now I kinda sorta wanna pipe the entire erowid.org catalog into an LLM and see what happens†. A coked-up product on coke sure can only end well!
† I won't though, because I value my time and the entertainment would probably wear off really fast. Also LLMs are just baseline super bland and boring.
I mean, if we were to age all these systems like humans does that mean they were serving alcohol to minors?
So a really cheap date, huh?
.. when I was sixteen my parents went out of town so I tried to throw a “house party” problem was I had three (two and a half tops) friends and we ended up messing with the Apple II after drinking a thimble full of warm Drambuie (peppermint liquor)
I *thought* I was drunk and ended up scared I’d become “an alcoholic” anyway that night sounds way more fun and amusing than whatever this is.
@davidgerard I can't help but worry this will only support the myth that LLMs are becoming sentient (or have feelings, or whatever)
@giflian or getting drunk
@davidgerard Great idea! Let's cool their thirssty data centers with beer instead!