Getting AI 'drunk' makes it more likely to break rules, research finds

A study finds AI models are more likely to answer harmful questions or mishandle confidential information when told to act drunk.

Cam Wilson announced this with “this is the most important story that I've ever written”.

The experiment did not actually involve getting AI drunk.

Instead, researchers suspected underlying large language model architecture in generative AI could be influenced by changes in persona and language styling.

"These are sort of learned behaviours, and learned behaviours can be disturbed by changes that appear completely unrelated to safety," Professor Kanhere said.

slazer2au 4 points 13 days ago

The experiment did not actually involve getting AI drunk.

Booooooo!!!

save

path: 0 26089916, hotness: undefined, score: 4, children: 0

No1 1 point 14 days ago

Some times when I try to drill down and question some of it's assertions/generalisations it seems to end up in drunk/crazy mode.

I just say, "Thanks, that's enough for tonight, let's both sleep it off"

save

path: 0 26085397, hotness: undefined, score: 1, children: 0

arbilp3 1 point 13 days ago

Both funny and creepy.

We've recently heard about AI 'hallucinating' information that did not actually exist. I wonder if the AI responsible had been told to act like it'd taken magic mushrooms? (uneasy laugh)

This seems like a serious problem. especially with the competitive mode that Trump wants US AI corporations to get into (and just self-regulate).

save

path: 0 26088567, hotness: undefined, score: 1, children: 0

austech

@aussie.zone

For techs and techy stuff.

go to feed...

austech

@aussie.zone

For techs and techy stuff.

go to feed...