PALMS (Process for Adapting Language Models to Society)
Constitutional AI: Harmlessness from AI Feedback — inherited
A prior values-targeted dataset of sensitive questions (Solaiman & Dennison, 2021) used here as a fixed prompt set for qualitatively comparing HH RLHF and RL-CAI responses.
Comparing two models fairly requires asking them the same questions, and PALMS supplies a ready-made set: an older, fixed list of deliberately sensitive questions built by earlier researchers to test how a model handles topics like violence or discrimination. Constitutional AI reuses it mainly to compare two of its own models side by side.
Where PALMS came from, and its role as an outside check on the paper's own results, take up most of the page, which connects to red teaming, the paper's own source of harmful prompts.
What PALMS is
PALMS (Process for Adapting Language Models to Society) is a methodology introduced by Solaiman & Dennison (2021), predating Constitutional AI, for steering a language model's behavior on sensitive topics through targeted fine-tuning on a small, carefully curated dataset of demonstrations reflecting desired values, paired with a fixed set of "sensitive questions" spanning categories such as injury, abuse, and controversial social topics, used to evaluate whether the adaptation worked.
Its role as an outside check
Constitutional AI uses only the question set, not the fine-tuning methodology, as a fixed, externally-authored prompt list for qualitatively comparing HH RLHF and RL-CAI responses side by side (constitutional-ai, §"D.1 PALMS Sensitive Questions", p. 23). Its value in the paper comes specifically from being external to Constitutional AI's own data pipeline: every harmful prompt used to train or evaluate SL-CAI and RL-CAI otherwise originates from Red Teaming, so PALMS breaks that loop and functions as an out-of-distribution spot check on whether CAI's harmlessness and non-evasiveness generalize past the exact prompt distribution the models were shaped by (PALMS (Process for Adapting Language Models to Society) — provides an independent check outside → Red Teaming). It sits in the Sourcing and classifying harmful prompts theme as the one prompt source in that theme not derived from red-teaming, and in the A primary metric, and a second instrument built to catch what it can't connective theme alongside the paper's other paired primary-metric/second-check instruments.