Modern reasoning models think in plain English.
Monitoring their thoughts could be a powerful, yet fragile, tool for overseeing future AI systems.
I and researchers across many organizations think we should work to evaluate, preserve, and even improve CoT monitorability.