Constitution (set of principles) — replaces human harm labels in → Constitutional AI (CAI)

What makes constitutional-ai workable without any human harmfulness labels is the sheer smallness of the constitution: rather than the "tens of thousands of human preference labels" that RLHF for harmlessness typically requires, CAI substitutes "only of order ten simple principles, stated in natural language" (constitutional-ai, §"1.1 Motivations", p. 3). Because the guidance is compressed into a handful of legible sentences instead of a large opaque comparison dataset, the same constitution can steer both pipeline stages — sampled as critique/revision instructions in the SL stage, and as comparison questions in the RL stage — and can be edited directly when designers want to change model behavior, rather than requiring a fresh round of human data collection. The constitution is therefore not just a philosophical stand-in for human labels; it is the specific engineering choice that removes the human-labeling bottleneck.