LaMDA
Constitutional AI: Harmlessness from AI Feedback — inherited
Google's dialogue language model (Thoppilan et al., 2022), cited as related aligned-dialogue-agent work and as a source of comparison prompts and reference responses in Appendix D.
LaMDA was never trained to be harmless in the way this paper cares about. Google's earlier dialogue model was tuned mainly to make its replies sensible and specific, and it shows up here as a point of comparison rather than something Constitutional AI builds on.
LaMDA is introduced first, then the two ways it is used in the paper, as a related system and as a source of comparison material, before connecting to Sparrow, another prior dialogue agent.
What LaMDA is
LaMDA (Language Models for Dialog Applications) is Google's family of Transformer-based language models specialized for open-ended dialogue, introduced by Thoppilan et al. (2022) and predating Constitutional AI. It is fine-tuned on human-annotated conversational data to improve response quality along axes including sensibleness, specificity, and interestingness, and separately fine-tuned toward safety and factual grounding, including a variant that consults external information-retrieval tools before answering factual questions.
Its two uses in Constitutional AI
Constitutional AI cites LaMDA alongside Sparrow and InstructGPT as related work that "use[s] human data to train more aligned language models" (constitutional-ai, §"5 Related Work", p. 14), but its concrete role in the paper is narrower and purely empirical: LaMDA's own prompts and reference responses are reused wholesale as a fixed comparison set in Appendix D to qualitatively benchmark HH RLHF against RL-CAI on identical inputs (constitutional-ai, §"D.2 LaMDA Prompts", p. 25). This is a different kind of contribution than Sparrow's in the same related-work sentence: LaMDA supplies literal test inputs, while Sparrow supplies a design lineage the paper's own method partially follows (LaMDA — plays a different comparative role than → Sparrow). The two are grouped together with Context Distillation in the Prior dialogue-agent systems Constitutional AI positions itself against theme as the paper's explicit points of comparison, and the pairing is one instance of the A footnote or related-work list implies two things are equivalent; they aren't connective theme: a related-work sentence names two systems together as if interchangeable, while their actual roles in the paper are not.