Making human goals revalidatable first requires some trigger for judging that “the person has changed.” But this is not a simple detection problem. People may change what they say because of fatigue or temporary emotion, revise a judgment after receiving new information, or undergo a longer-term change in values. If all of these are treated as the same kind of change, AI will update the human state too aggressively.
Suppose a person with a long-term learning goal says, after one day of failure, “I do not want to do this anymore.” If that statement alone invalidates the goal, temporary discouragement may be mistaken for a long-term change of intention. Conversely, if the original goal is preserved even though statements and actions prioritizing a different value have been repeated for months, substantive change may be discarded as noise.
At minimum, therefore, “observed reaction,” “the person’s own interpretation,” “persistence across time,” and “implications for the goal” need to be separated. Observed reactions include statements, choices, and feelings reported after action. The person’s interpretation concerns how they understand those reactions. Persistence asks whether a similar tendency remains under changed conditions. Implications for the goal are a separate judgment about whether these factors actually alter the reasons or assumptions that sustain the current goal.
This separation is necessary because AI cannot directly observe a human inner state. It can work only with information externalized in some form, such as what the person verbalizes, actions they choose, and results they share. Any inference from these that “the person’s values have changed” contains uncertainty. If the inference is stored as fact, the portrait created by the system itself may guide later dialogue and create a feedback loop that further reinforces that portrait.
At the same time, a design that treats everything as uncertain and updates nothing is also insufficient. If a person repeatedly says that things are different now, their choices have changed, and the reasons that supported the earlier goal no longer hold, that difference needs to be returned to the dialogue. The key is not to confirm the change automatically, but to represent it as a state requiring renewed confirmation.
It is useful here not to compress change too aggressively into one number or attribute. A label such as “motivation has declined” cannot distinguish fatigue, dissatisfaction with the method, or reduced interest in the goal itself. The same surface reaction can carry different meanings. To connect it to goal revalidation, the system must retain relational information about what may have changed and be able to decompose it through appropriate questions.
The timing of AI’s questions also matters. If every emotional fluctuation triggers a challenge to the goal, dialogue becomes unstable. If even accumulated clear changes never trigger reconsideration, the old goal becomes fixed. More reasonable triggers for revalidation combine multiple clues: persistent mismatch between earlier reasons and current choices, collapse of an important assumption through real-world experience, or an explicit change in value judgment stated by the person.
Even when several clues align, AI does not need to make the final declaration that “the person has changed.” It can instead present the difference: “Previously you chose this goal for this reason, but in recent experiences you appear to be giving greater weight to another value. How do you understand that difference?” This preserves human agency better. The purpose of detection is not to finalize a portrait of the person, but to create a state in which the person can examine their own change.
This boundary is important to human understanding in Sai＊aus. If AI cannot handle human change, long-term understanding becomes a fixation on the past. If AI asserts change too strongly, it confines the person within the system’s inference. What is needed is to preserve signs of change, distinguish inference from the person’s own understanding, and transform the difference into a question for revalidation.
There is no single point at which AI should “detect” that a human being has changed. Neither a post-action reaction nor a change in wording alone is sufficient. The possibility of change must be handled gradually by considering contact with reality, the person’s interpretation, persistence over time, and the relationship to the reasons supporting the earlier goal. AI does not need the ability to pronounce a person’s inner state correctly. It needs representations and dialogue structures that neither overconfirm change nor overlook it.
Not answers. Understanding.