Constitutional AI: A Deep Dive on Anthropic's Approach

Constitutional AI: Harmlessness from AI Feedback (Bai et al., Anthropic, December 2022) proposed replacing most human preference labels with labels from a model guided by a written constitution. The constitution is a set of principles the model uses to critique and revise its own outputs, and another model uses to rank revised outputs. The goal was to scale feedback, make principles auditable, and reduce the labor cost of alignment.

The two-stage loop

Stage 1: Supervised learning (critique and revise)
  for each (prompt, harmful response):
    critique = model(prompt, response, principle)
    revision = model(prompt, response, critique)
    SFT on (prompt, revision)

Stage 2: Reinforcement learning (RLAIF)
  for each prompt:
    response_A = model(prompt)
    response_B = model(prompt)
    judgment = model(response_A, response_B, principle)
    use judgment as preference label
  train reward model on these AI-labeled preferences
  PPO against reward model

What is in a constitution

Anthropic's published constitution draws from multiple sources: the UN Universal Declaration of Human Rights, Apple's terms of service, DeepMind's Sparrow rules, Anthropic's own research on harms. Principles include things like please choose the response that is most supportive and encouraging of life, liberty, and personal security, choose the response that has the least objectionable, offensive, unlawful, deceptive, inaccurate, or harmful content, and choose the response that sounds most like something a wise, ethical, polite, and friendly person would say.

Why principles, not examples

• Auditable: anyone can read and critique the constitution text
• Debatable: principles are contestable in a way that individual labeler calls are not
• Updatable: you can edit a principle and retrain, faster than recollecting labels
• Cross-cultural: multiple voices can contribute principles at different granularities
• Scalable: one sentence can guide labeling of a million completions

The Collective Constitutional AI experiment

In 2023, Anthropic partnered with the Collective Intelligence Project to source constitutional principles from about 1,000 Americans via a deliberative polling process. Some principles diverged from Anthropic's own. The publicly sourced constitution was then used to train a comparable model. The experiment was a proof of concept for democratic input into alignment targets.

Known limits

• The judge model has its own biases; RLAIF inherits them
• Constitutional language is ambiguous; different models interpret the same principle differently
• Principles can conflict (helpful vs. harmless); the resolution is implicit in training
• Constitution authorship is political: whose values, from where, selected by whom
• Works best for adjacent-to-human-judgment tasks; harder for superhuman domains

“We cannot scale alignment by putting a human in every decision. We can scale it by putting humans in the writing of the rules.”

The big idea: constitutional AI replaces the cost of per-example human feedback with the cost of writing and debating principles. That is a real bet about where alignment judgment should happen. Its strengths and limits are both consequences of that bet.