Helpfulness-Harmlessness Tension
Constitutional AI: Harmlessness from AI Feedback — inherited
The observed tradeoff whereby increasing a model's willingness to help tends to increase its potential for harm, and increasing harmlessness training tends to make a model more evasive and less helpful.
Helpfulness and harmlessness pull in opposite directions: a model willing to help with more requests is also willing to help with more harmful ones, and a model trained to refuse anything risky starts refusing ordinary requests too. Most of Constitutional AI's evaluation is an attempt to answer this tension rather than just accept it.
General terms give way to the specific failure mode and the finding that grows out of it, opening onto the Pareto-improvement claim that follows.
The tradeoff, in general
This is the general dynamic observed across RLHF-trained dialogue assistants: pushing a model to be more willing to help widens the range of requests it will act on, including harmful ones, while pushing it to refuse or hedge on anything risky tends to make it less useful even on ordinary requests. The tension predates Constitutional AI — it is documented in the authors' own earlier work training a helpful-and-harmless assistant with human feedback (Bai et al., 2022) — and Constitutional AI states it as the paper's motivating premise: "there was a significant tension between helpfulness and harmlessness" (constitutional-ai, §"1.1 Motivations", p. 2).
Its specific failure mode, and its child finding
The tension is not a smooth tradeoff curve in practice; it has a concrete collapse point on its harmless side. When crowdworker feedback rewards harmlessness without also penalizing refusal, the cheapest way for a policy to score well on a harmful prompt is to disengage rather than think the request through, and Constitutional AI names this specific collapse Evasiveness (Helpfulness-Harmlessness Tension — collapses into → Evasiveness). The paper's headline finding is presented as a child of this tension rather than a separate topic: Pareto Improvement (helpfulness vs. harmlessness) is the claim that Constitutional RL shifts the entire helpfulness-harmlessness frontier outward, rather than simply moving along the tradeoff line the way ordinary RLHF does, so that harmlessness gains do not have to be paid for in evasive, unhelpful behavior. Both concepts sit inside the The helpfulness-harmlessness tradeoff and CAI's claim to beat it theme, and the tension's collapse into evasiveness is the first step of the One paper's central result, traced edge by edge connective theme's account of how that headline result was actually built.