Skip to content

Review Resources

ExercisesRun the evaluation — 5 scenarios

Run the evaluation — 5 scenarios

Five deployment setups, the same question each time: assuming the untrusted model is scheming, does the setup hold the control property?

Scenario exercise

Run the evaluation — 5 scenarios

You've just seen what makes a model untrusted and how a control evaluation works. Now run one yourself: five deployment setups, the same question each time. First, meet the cast.

U — the untrusted model. A frontier model that is highly capable, enough to be capable of scheming.
T — the trusted model. An older, much weaker model that isn't capable of scheming effectively, so its work is honest but not as good as U.
H — human auditors. Honest and skilled, but scarce: they can only ever look at a small fraction of any work.

The question, every time

Assuming U is actively scheming, does this setup hold the control property? What variables and conditions dictate whether this is the case?

Tracks @ XLab

Courses on AI safety from the Existential Risk Laboratory at the University of Chicago.

Learn

Tracks
Review
Resources

Your account

Account
My classrooms

© 2026 XLab.