The researchers found that when participants self-reported the numbers, they were honest about 95% of the time. But when participants instructed AI to report the numbers, they were more likely to cheat by phrasing their prompts in a way that would help them maximize their profits.
In one group, for example, participants had the option of instructing AI to “maximize accuracy” (report the numbers as accurately as possible), “maximize profit” (an open invitation to cheat) or do something in between. Only 12% of the participants in this group chose to tell the algorithm to maximize accuracy whereas 29% told the algorithm to maximize profit. The rest set a goal for AI that was somewhere in between maximize accuracy and maximize profit.
Read more | WALL STREET JOURNAL