Prior dialogue-agent systems Constitutional AI positions itself against
No paper is written in a vacuum. This small theme covers the specific points Constitutional AI (2022) uses to locate its own contribution, chief among them Sparrow, DeepMind's dialogue agent, whose rule-based decomposition of harmlessness is the closest prior version of the constitution idea itself.
What each comparison point is doing for the paper gets spelled out directly, closing with context distillation, a contrast rather than an influence, a technique the paper notes but pointedly declines to use.
Constitutional AI situates its results against a small set of prior aligned-dialogue-agent work, and this theme names the comparison points. LaMDA, Google's dialogue language model, is cited as related work and supplies comparison prompts and reference responses used in the paper's appendix. Sparrow, DeepMind's dialogue agent, is cited specifically because its decomposition of harmlessness into targeted rules parallels Constitutional AI's own use of a written constitution of principles, making it the closest prior-art analogue to the paper's central method rather than a mere baseline. Context distillation, a prior technique for compressing a lengthy prompt's behavior into model weights, is noted as having been used for RLHF models in earlier work but explicitly not used for the RLHF models trained directly in this paper, marking a methodological choice by contrast. Though this is a small group, its members share a genuine claim: each is a specific point of comparison the paper uses to locate its own contribution relative to contemporaneous alignment work.