AI Alignment

The research and engineering work that aims to make an AI system pursue goals its developers and users actually intend. Alignment covers reward design, evaluation, training methods such as RLHF and DPO, and the broader question of whether a model's behavior is honest, helpful, and harmless. The field is closely tied to AI safety.

Definición editorial de DecipherU. Para definiciones técnicas autorizadas, consulta NIST AI RMF, ISO/IEC 42001, o las publicaciones oficiales de los laboratorios de IA.