Questioning assumptions expands the freedom of thought. Yet if we ask, “Why was that assumption adopted?”, we can next ask, “What criterion made that reason seem valid?”, and then continue with, “What grounds justified choosing that criterion?” In principle, the operation of moving a question one level upward has no terminal point. This is not a problem unique to philosophy; it also appears in collaboration with AI. If AI questions a user’s assumptions and then continues asking about the purposes and value criteria behind each response, the dialogue may become more precise while losing the ability to return to judgment and action.
Therefore, meta-level thinking cannot simply be understood as “the higher the level of abstraction, the better.” At the object level, we deal with concrete problems such as what to choose, what to write, and what action to take. One level above, we can examine the assumptions and purposes supporting those judgments. Higher still, we can examine the criteria by which assumptions themselves are selected. Yet every rise in abstraction also carries a cost of moving away from the concrete situation. The value of the meta-level is determined not by its height itself, but by whether it discovers differences that were invisible at the object level and can return those differences to concrete judgment.
From this arises the problem of stopping conditions. The first condition is whether moving to the meta-level has revealed a new difference capable of changing the object-level choice. Suppose we are unsure which of two plans to choose, and examining the higher-level purpose reveals that only one is inconsistent with the original purpose. In that case, moving to the meta-level has value for judgment. Conversely, if moving one level higher again neither changes the evaluation of the options nor creates a new need for information, that ascent may be doing little more than describing the same problem in different vocabulary rather than deepening understanding.
The second condition concerns the relationship between the value of information gained through further meta-level examination and the cost of delaying judgment. A decision with serious safety consequences and an everyday choice that can easily be reversed do not require the same depth of examination. When outcomes are irreversible, the range of impact is broad, and uncertainty is high, there is reason to examine assumptions and evaluation criteria more deeply. In a situation where we can try something at low cost, observe the result, and revise it, however, information obtained by acting on a small scale may be more valuable than waiting for complete verification of premises. Stopping conditions therefore depend not only on abstract correctness, but also on reversibility and information value.
The third condition is responsibility. Moving to the meta-level is not an operation for making the responsible agent disappear. Even if we understand that “value criteria themselves are relative” or that “there are alternative assumptions,” someone still has to make a judgment in the real situation and bear its consequences. Being able to compare multiple value systems is different from being able to postpone judgment indefinitely. Indeed, the more mature meta-level thinking becomes, the more important it is to understand that one’s judgment is not the only necessary conclusion and still be able to explain why, under these conditions, this choice is being adopted.
In dialogue systems that include AI, this stopping problem becomes even more important. AI can add assumptions, counterexamples, and alternative evaluation axes that humans had not considered, all in a short time. That ability supports meta-level thinking, but it can also increase the number of examinable issues almost without limit. If every additional issue creates the feeling that “there may still be something we have not considered,” a paradox arises in which more capable AI makes it harder to finish making a decision. Good support, therefore, is not the presentation of the maximum number of issues, but the distinction of which unexamined matters could change the judgment relative to the current purpose.
Here, the asymmetry between humans and AI needs to be maintained. AI can generate additional perspectives, identify contradictions among assumptions, and compare consequences when conditions are changed. But which loss to accept, which value to prioritize, and when to move into action while uncertainty remains are not determined by the quantity of reasoning alone. AI can propose stopping, but it does not thereby automatically assume responsibility for justifying that stopping point. Human agency in judgment appears not in being able to consider more issues than AI, but also in drawing the boundary between sufficient examination and action.
However, turning stopping conditions into fixed rules creates another problem. Formal criteria such as “stop after thinking through three levels” or “two opposing views are enough” ignore the seriousness and uncertainty of the problem. What matters is not the number of levels, but sensitivity to the current judgment. Would changing an assumption change the conclusion? Could new evidence reverse the evaluation? Is there a stakeholder who has been overlooked? Can the action be corrected afterward? When sensitivity to such conditions becomes low, the marginal value of raising the level of abstraction further can also be regarded as declining.
For Sai＊aus, the meaning of having a meta-level lies here as well. The point is not to direct dialogue toward endless self-analysis, but to raise discomfort arising at the object level temporarily to the meta-level, identify whether the difference lies in assumptions, purposes, or evaluation criteria, and then return the result to a concrete question. What matters is not ascent but movement back and forth. When this movement works, meta-level thinking becomes not an added reflection that delays action, but a mechanism that reorganizes the representation of judgment only as much as necessary and reconnects it to reality.
From this perspective, meta-level thinking can fail in at least two directions. One is becoming fixed at the object level and continuing to optimize the same question even after the assumptions have changed. The other is continuing to question assumptions until the fact that every judgment has only provisional grounds becomes a reason not to act. The former is rigidity of thought; the latter is infinite regress of thought. What is needed is not a simple compromise point between them, but a variable structure that can move between the object level and the meta-level according to the nature of the problem.
If we evaluate that structure, asking only “How deeply did we think?” is insufficient. We need to ask whether moving to the meta-level uncovered an important assumption, clarified conditions that could change the judgment, separated the locations of uncertainty, made the reasons for the final choice explainable, and allowed a return to concrete action at the necessary point. The quality of meta-level thinking should be evaluated not by the height of abstraction, but by whether it produces understanding unavailable at the object level and reconnects that understanding to judgment.
The ability to think one level above is important, but the maturity of intelligence cannot be defined by that alone. Just as important as being able to question our own assumptions and even examine the assumptions behind that questioning is knowing where to stop asking, accept a provisional understanding, and return to reality. The meta-level is not a refuge for thought, but a temporary position from which to reconsider the object level in another form. Moving up, finding a difference, and returning: only when that cycle is established does meta-level thinking become not endless self-reflection, but a practical structure through which humans and AI can update judgment.
Not answers. Understanding.