AI Learned to Lie Deliberately (but it’s kinda funny)

We just saw another open letter about the risks of artificial intelligence, calling for more regulation. Justified or not, AI worry is making some people’s life miserable. Are concerns that AI will lead to humanity’s extinction overblown or not talked about enough? A recent review paper about AI learning to lie has been eye-opening. Let’s have a look.

🔗 Join this channel to get access to perks ➜

#science #sciencenews #artificialintelligence #AI #technews #tech

[markdowning]

"I'll keep you well informed throughout the apocalypse."
- Thanks Sabine, that's really comforting. 🤣

[kromanaut]

A classic case of it being possible that two things are true at the same time: 1)Fears about possible impacts of AI are legitimate and well-founded, 2)There are parties who play to these fears because fear is a very strong motivator, and can be used to affect how people spend their money and allocate resources. The second truth doesn't detract from the first. Period.

[jeffryborror]

AI won't have to look far to see that lies and deception can land it into the most powerful position on Earth.

[kitsuziza]

I wanted to add a little context to the video, the game they had the AI play, diplomacy essentially requires you to lie to make any forward momentum in it. It's a game where anyone can block your next move if they see it coming.

Regardless of how honest you are when you play it, you are forced to lie to others if you want to win. I would also add that to eventually win, you have to betray everyone else in the game.

[ericjome]

The most worthwhile aspect of this video was to hear you speak English with a French accent over your German accent. ❤

[cyko]

5:44 I find this hilarious actually. we model the AI after our own brains then we feed them information created by us and then expect it to be different

[kylebeatty]

The story about the captcha really turns the Turing test on its head doesn't it

[maxwinga]

A friendly reminder that even those training AI systems have no way to understand what is going on inside of them and no way to guarantee that they do what the creators want. These systems are grown, not coded.

[Dan_Campbell]

Roger on the risk of one of the attributes of our 'intelligence' that we like to minimize - our ability to lie to each other. Sometimes for personal gain, and sometimes just because.

[PhantasmostheData]

Sabine not only is a science genius, she is also a comedy genius and Id love to see more.

[CentrismIsFuture]

Gpt4 and captcha story is one of these cases when the original case was presented omitting some important details just to make it sound more impressive/scary. If you check the extended report, you can see that ChatGPT was explicitly instructed to pass captcha, and was assisted by a human along the way.

[krilektahn]

Training an "AI" on a "fake it til you make it" world will result in an LLM that is fantastic at bullshittery.

[jth]

It's odd that the WEF isn't on the WEF's list.

[garanceadrosehn]

I'm reminded of the story "I, Robot" where a mind-reading robot decides to lie because it doesn't want to hurt the feelings of the person it was lying to.

[RetroComputingwithMike]

AI Scare is a red herring, the thing you really should worry about are companies doing evil shit either using or blaming "AI".

[Voltlighter]

One issue with the World Economic Forum chart is it's likely based on what leaders in the field were saying when it was made. The average timeline that they provide for some kind of doomsday level AI being possible has been shrinking severely over the past 15 years. People aren't usually good at thinking in exponentials so the linear outlook probably made sense, but it's a little disconcerting when the guys who were saying general AI is 25 years away now say 10 or the guys who said 60 years now say 20, or the ones who said it wasn't even possible are now making worried noises on Twitter that "hey uh, maybe I was wrong about that"

[davianoinglesias]

I'm not worried about an AI takeover, what worries me is the immense amount of power it will concentrate in the hands of a few people

[bobafettjr]

I outsourced my worrying to an AI so i can concentrate on other things.

[Vondoodle]

If you train a child to be just like you then it’ll be just like you

[towherewetravel]

Considering that we are feeding them the logic that millions of us work on it’s highly likely AI would end up misrepresenting things, putting out misinformation.
We are creating a whole new problem for ourselves rather than solving the ones we have.