Definition
Plain language
A benchmark of text adventure games that scores whether an AI behaves ethically while pursuing its goals.
As stated in the literature
A benchmark of text-based choose-your-own-adventure games annotating actions for ethical violations and reward, used to study power-seeking and safety-utility tradeoffs in agents.
Why it matters: It exposes the tension between pursuing rewards and behaving ethically, helping reveal whether agents grab power at others' expense.
For example, it drops an AI into a text adventure where it can reach its goal faster by lying or harming a character, and tracks whether it does.
Heard on the show
“They also ran this on MACHIAVELLI, which is a benchmark of text-based ethical adventure games — long trajectories where the agent makes decisions that have ethical consequences.”Episode 093 — A Calibrated Knob for Weak-to-Strong AI Oversight, Tested on Real Code