AnthropicBased on real interviews
Constitutional AI failure modes
Investigate how constitutional principles and AI-generated feedback can produce failures during training.
Unlock with Full AccessThe challenge
Your research lead wants an account of what can go wrong inside a Constitutional AI training loop. Choose specific failure modes, explain how the training process produces them and propose evidence that could distinguish your explanation from alternatives. Finish with an experiment worth pursuing over a three-month research window.
What you'll practise
Training mechanisms
Connect a failure to the principles, critique process or preference signal that produces it.
Discriminating evidence
Propose evaluations that test your explanation, including what would change your view.
Research priorities
Choose a tractable investigation and explain the trade-offs of your proposed mitigation.
Write your reasoning and submit it for feedback.