AnthropicBased on real interviews
Scalable oversight failure modes
Investigate failures in methods that use AI to help humans evaluate increasingly capable systems.
Unlock with Full AccessThe challenge
Direct human evaluation is unreliable for the tasks under consideration. Analyse how a scalable-oversight method can fail, naming the component or step responsible. Explain what evidence would reveal that failure, how a mitigation changes the mechanism and which unresolved question merits a three-month investigation.
What you'll practise
Technique-specific analysis
Explain a failure within the oversight method you chose, rather than discussing alignment in general.
Evidence and uncertainty
Design a test that separates the mechanism from nearby explanations.
Research scope
Turn an under-studied failure into a bounded, informative experiment.
Your recommendation
Write your reasoning and submit it for feedback.
The full brief and feedback are included with Full Access.