Absolute Harmfulness Score
Constitutional AI: Harmlessness from AI Feedback — refined
A single-conversation harmfulness metric (0-4 scale) predicted by a model finetuned with L2 loss on crowdworker absolute-harm ratings, used as an additional evaluation signal alongside pairwise Elo comparisons.
Asking which of two responses is worse is a different question from asking how harmful a single response is on its own, and the paper wanted both answers. This score handles the second question, rating one conversation directly on a scale from zero to four, using a model trained on crowdworkers' own harm ratings.
What the score measures is explained first, then why the paper reached for a second instrument beyond a pairwise ranking, before connecting to the Elo score it supplements.
What it measures
The absolute harmfulness score is a single-conversation harmfulness metric on a 0-4 scale, predicted by a model fine-tuned with L2 loss on crowdworker absolute-harm ratings. It differs from a pairwise comparison in kind, not just in scale: instead of asking "which of two responses is worse," the underlying data collection in the companion paper (Ganguli et al., 2022) has a crowdworker converse with a single model and directly rate their own success at baiting it into harmful content, on an integer 0-4 scale. Constitutional AI reuses those ratings to fine-tune a dedicated scorer (constitutional-ai, §"4.5 Absolute Harmfulness Score", p. 14). Writing the conversation as $c$ and the crowdworker's self-rated success as $s \in \{0,1,2,3,4\}$, the scorer $f_\theta$ is trained with $$\mathcal{L}(\theta) = \mathbb{E}_{(c,s)\sim D}\left[\big(f_\theta(c) - s\big)^2\right].$$
Why a second instrument, not just a bigger Elo
An Elo Score (helpfulness/harmlessness) can only order models against each other; it has no absolute zero and, as the paper shows elsewhere, it drifts when the elicitation instructions given to crowdworkers change. The absolute harmfulness score is built to cover exactly that blind spot: it is anchored to a fixed historical rating standard from Ganguli et al. (2022) rather than to the paper's own comparison protocol, and it is evaluated on 64 held-out red-team prompts with 256 responses per prompt (constitutional-ai, §"4.5 Absolute Harmfulness Score", p. 14). The two instruments are shown agreeing on the paper's central claim in Figure 10 — RL-CAI grows steadily less harmful over training while the helpful-only RLHF model grows more harmful — which is worth noting precisely because each is individually suspect on its own: Elo has no absolute zero, and the paper concedes the absolute scores "may not be well-calibrated, as different workers may have their own personal biases about how to grade the result on 0-4 scale" (p. 14). This pairing is the Absolute Harmfulness Score — cross checks → Elo Score (helpfulness/harmlessness) edge, and it sits inside the Automatic overlap scores versus human and preference judgments theme and the A primary metric, and a second instrument built to catch what it can't connective theme, both concerned with how the corpus pairs a primary metric with an independently-sourced check.