OpenAIBased on real interviews
Deliberative alignment failure modes
Examine where reasoning over a written safety specification can fail to constrain a model's answer.
Unlock with Full AccessThe challenge
The model is trained to recall and reason about safety specifications before responding. Your memo must explain failures specific to that process, from how the specification is represented to how the final response is produced. Propose ways to test the mechanisms you identify and select a focused research direction.
What you'll practise
Specification reasoning
Trace how a failure arises within the training or inference process.
Mechanism testing
Separate a plausible explanation from one supported by a discriminating evaluation.
Mitigation trade-offs
Explain what your proposed change fixes and which risks remain.
Your recommendation
Write your reasoning and submit it for feedback.
The full brief and feedback are included with Full Access.