chapter eight

8 Alignment through principles: Constitutional AI

 

This chapter covers

  • Reinforcement learning from AI feedback guided by written principles
  • The helpfulness-honesty-harmlessness trade-off
  • Designing a constitution, from Anthropic’s principle list to OpenAI’s Model Spec
  • Teaching a model the spec through midtraining

Reinforcement learning from human feedback has a scaling problem. Human labelers are the bottleneck for reward modeling. InstructGPT used about 13,000 SFT prompts, 33,000 reward-model prompts, and 31,000 PPO prompts. By post-training standards, that is a small dataset. Once a team needs to update its assistant for a new product surface, policy area, or to fix misalignments that emerged in production, the cost of collecting another round of labeled comparisons starts to dominate. The hallucination examples from the last chapter are not a one-time fix. Every new earnings cycle or out-of-training distribution data introduces new transcripts, entities, and ways for financial research assistants to fabricate plausible-sounding names. RLHF can train against the failures the labelers have already seen, but the human labeler workforce cannot scale to cover every variant in a real product distribution.

8.1 The case for principles

8.1.1 Where RLHF runs out of road

8.1.2 Helpfulness, honesty, and harmlessness

8.1.3 Principle versus preferences

8.1.4 The three-pillar lens on principle-based alignment

8.2 Constitutional AI

8.2.1 The two-phase pipeline

8.2.2 The constitution

8.2.3 Supervised learning from critique and revision

8.2.4 Reinforcement learning from AI feedback (RLAIF)

8.2.5 Why this matters

8.3 What makes the AI labeler reliable

8.3.1 RLAIF as the training mechanic

8.3.2 Calibrating the AI Judge

8.3.3 Failure modes when the constitution is the only signal

8.4 OpenAI’s Model Spec

8.4.1 From Anthropic’s prose constitution to a published spec

8.4.2 The chain of command

8.4.3 Mapping the spec to the three pillars

8.4.4 Testable published spec

8.5 Writing and aligning to your own spec

8.5.1 Drafting a spec for the financial research assistant

8.5.2 Distilling the spec through critique and revise

8.5.3 Model spec midtraining

8.5.4 Operating an in-house spec

8.5.5 Limits, trade-offs, and where Constitutional AI fits in alignment

8.6 Summary