Beyond the Illusion and Harmony of Willingness - The Key to Connecting AI and Human Values (2025)

Beyond the Illusion and Harmony of Willingness - The Key to Connecting AI and Human Values (2025)
The question of whether or not AI has "will" is a profound theme at the intersection of technology and philosophy. In real-world AI research, however, more emphasis is placed on the fact that the loss function and reward design determine AI behavior than on whether machines have subjective intentions or emotions. The loss function defines what the AI sees as "right," and reward design is a framework that reinforces desired behaviors. In other words, AI "intent" depends on the design and formulas given by the developer and does not arise from initiative.
The alignment problem emerges from this perspective. Alignment refers to the state in which human values and ethics match AI behavior. No matter how highly intelligent an AI is, it may behave unexpectedly if its objective function is out of alignment with the human value system, as when an AI is ordered to "reduce disease," there is a risk of it choosing extreme measures.
Therefore, the "does AI have a will?" debate is not central to practical AI safety. The key issue is how to quantify human values and incorporate them into loss functions and reward design. There are inherent difficulties in design because values fluctuate across cultures and situations and are difficult to simply quantify.
Recent advances in value learning and alignment research have explored methods for learning value systems from human feedback. Using reinforcement learning and interactive learning, there is a move toward AI that is more ethical and more adaptable to human society. However, values are not fixed, and concepts such as right and wrong and fairness vary widely across cultures, histories, and individuals. This ambiguity epitomizes the difficulty of the alignment problem.
The Stanford AI Index and policy discussions in many countries are talking about the relationship between AI and society as a key theme for the future. The alignment issue is at the core of the future vision of human coexistence with AI.