Sparrow
Constitutional AI: Harmlessness from AI Feedback — inherited
DeepMind's dialogue agent (Glaese et al., 2022), cited as related work whose decomposition of harmlessness into targeted rules parallels this paper's use of constitutional principles.
Constitutional AI is not the first system to try breaking harmlessness down into a set of written rules. Sparrow, DeepMind's earlier dialogue agent, did something close to the same thing, trained with human preference feedback and able to back up factual claims with live search results.
Sparrow is introduced first, then why the paper treats it as an ancestor in its own design lineage rather than just another baseline to beat, before connecting to LaMDA.
What Sparrow is
Sparrow is DeepMind's dialogue agent, introduced by Glaese et al. (2022), "Improving Alignment of Dialogue Agents via Targeted Human Judgements," and predates Constitutional AI. It is trained with reinforcement learning from human preference judgments and is grounded with live search-engine queries to support factual claims with evidence, and its harmlessness training is organized around roughly two dozen hand-specified rules targeting particular failure categories — for example, rules against offering specific financial, medical, or legal advice, or claiming to have a body or to feel emotions.
Why Constitutional AI treats it as design lineage, not a mere baseline
Constitutional AI's related-work section lists Sparrow alongside LaMDA and InstructGPT as prior work that "use[s] human data to train more aligned language models" (constitutional-ai, §"5 Related Work", p. 14), a phrasing that makes the three sound interchangeable. Sparrow's actual role is narrower and more specific: the paper singles out "Sparrow's decomposition of harmlessness into different areas" as having "some commonality with our use of principles forming a 'constitution'" (constitutional-ai, §"5 Related Work", p. 14-15), making it the closest prior-art precedent for Constitutional AI's central design idea — targeted written rules standing in for a mass of individual human labels — rather than just another comparison point. This is a genuinely different contribution than LaMDA's in the same sentence, which supplies literal test prompts rather than a design lineage (LaMDA — plays a different comparative role than → Sparrow). Both sit inside the Prior dialogue-agent systems Constitutional AI positions itself against theme, and their pairing in one related-work sentence is an instance of the A footnote or related-work list implies two things are equivalent; they aren't connective theme's broader pattern: a related-work list implies an equivalence that a closer reading dispels.