Volume 1,Issue 2
Value alignment, an effective means and an unavoidable stage in the sound development of artificial intelligence (AI) technology, aims to bring the capacities and behaviour of large models into conformity with genuine human intentions, values and social moral norms. Existing technical paths and ethical approaches to AI value alignment are feasible, yet they carry limitations and shortcomings. The technical path to value alignment lacks effectiveness and scalability and is constrained by subjective human preferences; the weak-approach path to value alignment suffers from the "alignment gap", the difficulty of unifying values, and the problem of static values; and the strong-approach path to value alignment confronts such difficulties as the irreducibility of morality to a capacity, the predicament of affective computing technology, and the complexity of aligning multiple agents. The interactive value alignment path is an effective route to realising AI value alignment: granting AI interactive agency is the precondition of value alignment, contextualised value consensus is the key to interactive value alignment, and value alignment is realised through human-machine cooperation and social scenario simulation.