Argues that developers should evaluate models both for dangerous capabilities and for their propensity to use them, and that these evaluations should feed into training, deployment and governance decisions.
This is an external publication hosted at arxiv.org. AI Governance LK did not write it; the summary and notes on this page are our own.
Why it matters
The paper connects the technical practice of evaluation to the governance question of what developers should do with the results — a link that matters for anyone drafting policy.
What you will learn
- Distinguish dangerous-capability from alignment evaluations
- Explain where evaluations sit in a responsible deployment process
Before you start
Suggested background. This is editorial guidance, not a requirement.
- None
How to cite
Shevlane, T., Farquhar, S., Garfinkel, B., et al. (2023). Model Evaluation for Extreme Risks. arXiv:2305.15324.