In problem solving with AI, local improvements can proceed very quickly. Goals such as shortening text, reducing cost, decreasing processing time, or ranking candidates are relatively easy to state explicitly and can be optimized iteratively. Yet improvement of a local objective does not guarantee movement toward a higher-level purpose. If the initial purpose was “reduce the user’s burden,” but evaluation gradually focuses only on “minimize the number of operations,” the number of operations may decrease while difficulty of understanding increases.
To detect this problem, it is not enough simply to preserve the purpose as a sentence from the beginning of the dialogue. The higher-level purpose, current question, active constraints, evaluation metrics, and generated output need to be kept distinct in terms of how they relate to one another. Which part of the higher-level purpose does the current question address? Is the evaluation metric a valid proxy for the purpose? Has a constraint added along the way altered the purpose? Drift in purpose appears less as an error in any single element than as a change in the relationships among elements.
However, not every departure from the initial purpose should be judged abnormal. New facts may emerge through dialogue, and the user may legitimately change the purpose. For example, a plan may begin with the purpose “travel in the shortest possible time,” but after new safety information is learned, it may change to “prioritize safety even if it takes somewhat longer.” In this case there is a difference from the initial purpose, but it is not a failure; it is the result of an update in understanding.
What should therefore be detected is not simple difference, but the path by which the purpose changed. We need to distinguish whether the change was explicitly approved, justified by new information, or implicitly substituted while pursuing an easily measurable metric. AI can show differences from earlier dialogue and point out, “Safety was originally a condition, but the current comparison uses only cost.” Whether that change is acceptable, however, remains for the user to judge.
A mechanism for detecting drift in purpose also faces the problem of false alarms. Human purposes cannot always be expressed in a single clear sentence, and multiple values may exist in tension. If AI warns about every slight difference in wording as a change in purpose, the dialogue becomes dominated by confirmation. If sensitivity is too low, important changes in value criteria may be missed. What is needed is an emphasis not on matching words, but on differences capable of changing the judgment outcome.
As an implementation principle, it is useful to trace the relationship between purpose and means backward at regular intervals. Instead of asking only, “Does this output answer the current question?”, we return to “Which purpose does this question contribute to?” and “What grounds support the expectation that improving this evaluation criterion will also improve the higher-level purpose?” This backward check allows evaluation of output quality to be separated from evaluation of fit with purpose.
There is room here for human meta-level thinking and AI support to connect. Humans bear the meaning and value of purposes, but over a long dialogue they may forget earlier conditions or have their attention captured by the number in front of them. AI can supplement such oversights by comparing the history and externalizing changed conditions. At the same time, AI itself adapts strongly to the current context, so leaving preservation of purpose entirely to AI is not necessarily sufficient. Purpose checking needs to be designed as a distinct operation.
The ability to detect drift in purpose is not the ability of AI to know the “true purpose.” Human purposes may be formed through dialogue, and conflicting purposes may coexist. What AI can do, therefore, is show inconsistencies that arise among declared purposes, current questions, evaluation criteria, and outputs, and prompt reconsideration. What matters is not inferring and fixing the purpose, but making the possibility that the purpose has changed available for human examination.
From this perspective, more capable AI does not make purpose checking unnecessary. On the contrary, the faster local optimization becomes and the more persuasive outputs can be generated at scale, the harder directional drift can be to notice. Rather than weakening AI’s capability for safety, we need another layer in which the direction of that capability can be examined. Design should circulate both speed and directional checking instead of treating them as opposites.
Asking “what for?” is not merely a questioning technique. When humans and AI explore together, it is an operation for checking whether the current question remains connected to the higher-level purpose. Rather than allowing purpose, question, and output to flow only in one direction, we need a path back from output to question and from question to purpose. With that cycle, AI can not only accelerate methods, but also support humans in discovering whether the methods have left the purpose behind.
Not answers. Understanding.