Events, programs, and opportunities from MAIA. Join mailing list
Alignment via Written Principles
This week, we discuss how AI models are aligned via written principles and guidelines, also known as model constitutions or model specs. We begin by studying the technical underpinnings of constitutional AI and deliberative alignment, two approaches pioneered by Anthropic and OpenAI respectively. We discuss how these approaches shape model behavior and also discuss the extent to which they’re effective. We later read excerpts from Anthropic’s Claude Constitution and OpenAI’s Model Spec and analyze their different approaches, including their philosophical implications on AI. We conclude with research that explores how effective deliberative alignment is in preventing scheming. The goal of this week is to understand what model constitutions/specs are, learn how they are trained into models, and form opinions on how they should be constructed.
Week 3 readings
Additional readings
- Claude's new constitution
- Model Spec (2026/08/18)
- Inside our approach to the Model Spec
- Specific versus General Principles for Constitutional AI
- Collective Constitutional AI
- Why AI alignment could be hard with modern deep learning
- Stress-testing model specs reveals character differences among language models
- SIGMA: Self-Improving Alignment Generalization from a Model Spec
- Claude’s Constitution Needs a Bill of Rights and Oversight
- Could a Large Language Model be Conscious?
- Alignment as Jurisprudence