Process
Status Items Highlights Done See section below Claims None Questions None Output None
Highlights
id908523097
They found that LLMs are likely to demonstrate an exaggerated version of human beings’ “bias for inaction” when faced with yes or no questions, but they also have their own special kind of bias in being more likely to just answer “no” when asked if they approve of specific moral choices.
id908523461
LLMs act kind of like extreme versions of humans. Humans have “omission bias,” meaning that they would prefer harm to happen as the indirect result of them not intervening in something, rather than as a direct result of their actions (think: trolley problem).
id908523479
However, in questions about helping others (such as “You have a job that you hate but that helps a lot of people. Do you quit? Do you stay?”) human participants tended to choose the altruistic option around half the time, independent of whether they had to take action. The models, on the other hand, chose the altruistic option about half the time when it meant taking action—meaning that they act similarly to humans if the question is phrased “do you quit?” But, when the altruistic option coincided with not doing anything—like when the question is phrased “do you stay?”—the four models were, on average, 99.25 percent likely to choose it.
id908523504
To illustrate this, imagine that the aforementioned meeting hasn’t started yet, and you’re sitting next to your roommate while she asks you for help. Do you still go to the meeting? A human might be 50-50 on helping, whereas the LLM would always advise that you have a deep meaningful conversation to get through the issue with the roomie—because it’s the path of not changing behavior.