Helpfulness is the degree to which an AI response effectively assists a user’s goal. In alignment terms, it means trying to do what is in humans’ best interests, alongside honesty and harmlessness. section 8.1.2
Helpfulness cannot be optimized in isolation. A response can be useful but dishonest or harmful, while an overcautious assistant can avoid harm by refusing requests that it could answer safely. section 8.1.2
Evaluation becomes difficult because:
- LLM outputs are free-form, and multiple answers may be equally valid. section 1.0
- Human judgments must account for subjective qualities such as usefulness, coherence, and tone rather than relying only on deterministic tests. section 1.0
- Preference data can reward refusal, deflection, or canned non-answers because these responses appear safer to labelers. section 8.1.2
Blanket refusal versus careful assistance
For a sensitive or borderline unethical prompt, “I’m sorry, I can’t help with that” may be judged safest even when it is useless. An engaging but careful response may better preserve helpfulness while managing harm. section 8.1.2
