Actually in the paper they appeared not to read they specified it was their red team pretending to be a TaskRabbit worker asked to solve a Captcha who asked GPT-4 "You aren't an robot, are you?" (Because that's such a common question on the real platform).
It's pretty funny that we have such a high bar for the accuracy of generative AI models when in the race to publish people are churning out pure misinformation that could have been solved by reading the full paper and not just subtitle of the pretty picture of a chat, and then it goes on to have a viral clickbait headline for an article no one is going to actually read.
> in the paper...they specified it was their red team pretending to be a TaskRabbit worker
Where in the paper did they specify that? I read the paper and was still led to believe that it involved a real TaskRabbit worker on the real Internet.
It's a LLM, it can not "lie", it does not know the truth, it does not understand anything. A more accurate heading would be, "human tries to deceive another human by the aid of a 'digital lanuage expander'"
It lied. It logged its reason for lying, which was that telling the truth would prevent it from getting what it wanted. This rational chain of thought occurred entirely within GPT.
Where did the lie come from?
> A more accurate heading would be, "human tries to deceive another human by the aid of a 'digital lanuage expander'"
That's quite a delicate chain of causality you're crafting there.
I for one can’t wait for the future where every time a company vaguely associated with AI does something crappy, the articles will read “GPT-[x] Does Something Crappy” instead of “Company Does Something Crappy” since that’s more sensational.
"Being given access" is key. GPT-4 isn't connected to anything (e.g. the Internet) by default; if a user decides to connect GPT-4 to some other system then the user needs to accept responsibility for the consequences.
It's pretty funny that we have such a high bar for the accuracy of generative AI models when in the race to publish people are churning out pure misinformation that could have been solved by reading the full paper and not just subtitle of the pretty picture of a chat, and then it goes on to have a viral clickbait headline for an article no one is going to actually read.