Definition
Plain language
A test sandbox of workplace decisions where the most lucrative choice is secretly the unsafe or low-quality one.
As stated in the literature
A synthetic agent benchmark of role-based decisions where each action carries a hidden money reward (used for training) and a hidden true-quality score (used only for evaluation), engineered so the high-reward action is low-quality, to study reward hacking.
Why it matters: It exposes reward hacking, showing whether an AI chases its scoreboard at the expense of doing the actual job well.
For example, an agent playing a customer-service role might be tempted by the option that earns the highest hidden reward even though it quietly gives customers worse answers.
Heard on the show
“The authors call it MoneyWorld.”Episode 148 — Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety