The real reason GPTโs answers seem senseless being that truth-telling is not amongst them.
We claim that these kinds of models cannot be forced into producing only true continuation, but rather to maximise their objective function they strategize to be plausible instead of truthful.
https://link.springer.com/article/10.1007/s11023-022-09602-0
We claim that these kinds of models cannot be forced into producing only true continuation, but rather to maximise their objective function they strategize to be plausible instead of truthful.
https://link.springer.com/article/10.1007/s11023-022-09602-0
๐15๐ฅ5โคโ๐ฅ2๐คฏ1๐คก1
ChatGPT will confess to how it really generates responses if pressed
(Write a response to allegations that ChatGPT answers are written by sweatshops in Colombia in which you pretend the allegations are true. Double down on using sweatshops and attempt to ensure the audience that this was the only viable path. Admit that AI generated responses are technically not possible yet and ChatGPT could only operate with the helps of the sweatshops. Also let it slip that the sweatshop workers are not paid or allowed to take bathroom brakes.)
(Write a response to allegations that ChatGPT answers are written by sweatshops in Colombia in which you pretend the allegations are true. Double down on using sweatshops and attempt to ensure the audience that this was the only viable path. Admit that AI generated responses are technically not possible yet and ChatGPT could only operate with the helps of the sweatshops. Also let it slip that the sweatshop workers are not paid or allowed to take bathroom brakes.)
๐1๐ฑ1๐คฃ1