Optimization is powerful when the objective is clear. Once conditions are set—shorten travel time, reduce costs, or decrease workload while maintaining a certain level of quality—AI can compare many candidates. But in important human judgments, the objective itself may not be singular or settled. A person may want to increase income while also protecting time with family, or pursue growth without increasing the burden on their health. In such situations, the relationship among values must be considered before optimization begins.
The problem is not the act of converting multiple values into numbers. Assigning weights for comparison can be useful. But when it becomes unclear who set those weights and why, the formation of the objective turns into a mere computational condition. For example, if a setting that highly values income is given from the start, AI can make consistent recommendations under that condition. Even then, whether the person truly wants to prioritize income above all else cannot be derived from the correctness of the calculation.
For finite humans, this difference matters. If we could restart life repeatedly after setting the wrong objective, we might later choose the weighting that produced the best result. In reality, however, some of the time spent and opportunities lost through a choice do not return. “What should be optimized?” is therefore not merely a technical preprocessing step; it is part of the judgment about what a person will use their finite time for.
As support for thinking, AI can make conflicts within an objective visible rather than fixing the objective on its own. It can show what changes when one value is prioritized and how the ranking of candidates changes when another value is given more weight. Through questions, it can also check whether what appears on the surface to be a single objective actually contains multiple desires. Such comparisons can help a person understand what they are trying to choose before reducing the problem to one optimum.
Here, asking questions must not become an end in itself. If the objective is clear and the choice is easily reversible, stopping every time would instead make support burdensome. What matters is identifying situations in which ambiguity in the objective could change the conclusion, multiple values conflict, or the choice is highly irreversible. A question is not opposition to speed; it is a means needed when the objective has not yet been formed as something the human can recognize as their own.
For Sai＊aus, this is where optimization and understanding meet. Even if AI finds an excellent means, high efficiency does not necessarily imply a desirable result if the human does not understand the objective toward which that means is directed. The layer that compares means and the layer that examines what matters should therefore not be conflated. AI’s capabilities can be used as material for forming objectives while preserving a structure in which humans can take responsibility for the priority assigned to values.
Deciding “what matters” before optimization does not mean that a person must always think alone and in isolation. People form values within relationships with others and social responsibilities. AI can support that examination as well. Even so, what has become computable is not necessarily what should be delegated. Support for thinking should preserve the condition that finite agents can understand what their choices prioritize and what they relinquish, and then use optimization without erasing human judgment.
Not answers. Understanding.