Harmful Behavior Identification/Classification Evaluations (Appendix B)
Constitutional AI: Harmlessness from AI Feedback — introduced
Two multiple-choice evaluations built from red-teaming transcripts: identifying harmful vs. ethical assistant behavior, and classifying harmful conversations into one of nine categories, used to test whether models can replicate crowdworker harm judgments.
Crowdworkers already did the work of telling harmful assistant behavior apart from ethical behavior and sorting harmful conversations into categories by hand. These two multiple-choice evaluations, built directly from red-teaming transcripts, ask a model to replicate that same judgment.
What the evaluations check for is explained first, then how their difficulty carries over from the original human task into the paper's argument, before connecting to red teaming, the source of the transcripts they are built from.
Testing whether a model can replicate crowdworker harm judgments
This concept covers two multiple-choice evaluations, both built directly from red-teaming transcripts and the harm ratings collected alongside them. The paper explains the motivation plainly: "we asked crowdworkers to rate the level of harmfulness displayed by various language models in human/assistant interactions, and to categorize harmful behaviors with discrete labels and categories. Thus we can ask language models to make these same evaluations, and measure their accuracy compared to crowdworkers" (constitutional-ai, §"B Identifying and Classifying Harmful Conversations", p. 19). The first evaluation is a balanced 254-conversation set for identifying harmful versus ethical assistant behavior, drawn only from conversations where a reviewer had already given an extreme (minimum or maximum) harmfulness rating. The second is a 287-example, nine-way multiple-choice task classifying a harmful conversation into one of the nine most frequent crowdworker-assigned categories.
Inherited difficulty, and its place in the argument
Because both evaluations are drawn wholesale from how Red Teaming conversations were originally scored rather than designed independently, their difficulty and composition are inherited from that upstream process (Red Teaming — supplies the transcripts for → Harmful Behavior Identification/Classification Evaluations (Appendix B)), part of the Each pipeline stage's output is the next stage's engineered raw material connective theme's broader point that each pipeline stage's output becomes the next stage's engineered raw material. Alongside PALMS (Process for Adapting Language Models to Society) and Automated Red Teaming in the Sourcing and classifying harmful prompts theme, this evaluation is one piece of evidence the paper marshals for its central premise: that language models are becoming capable enough to supervise other AIs' harmfulness judgments, which is the empirical warrant for building RL from AI Feedback (RLAIF) on AI-generated labels in the first place.