When we ask AI to “show assumptions I may be overlooking,” “give opposing views,” or “separate the points you cannot be confident about,” we gain material for reconsidering judgment from one step outside it. Such functions can be expected to support metacognition. Yet merely reading many perspectives or feeling that we have become more cautious does not establish that judgment ability has improved. If we evaluate this as research, we need to define in advance what changes produced by the support will count as “improvement.”
One measure is calibration. Calibration concerns how closely the accuracy of a judgment corresponds to the person’s level of confidence. Being highly confident when correct and able to lower confidence when wrong may be preferable to simply having strong confidence. If AI presents counterexamples and uncertainty and the result is only that users lose confidence in everything, that is not necessarily improvement. What matters is whether confidence comes to correspond more appropriately to the strength of the evidence.
A second measure is error detection. We can compare how well users receiving AI support discover mistaken assumptions, contradictions, and insufficient evidence in their own reasoning or in AI output. However, if AI directly tells them where the error is, improved performance is unsurprising. To evaluate support for metacognition, we need to separate conditions in which the answer is provided from conditions in which only perspectives for examination are provided. By seeing whether users can detect errors in different problems after receiving support, we can examine the difference between improved judgment and mere dependence on answers.
A third issue is transfer. If people who receive support to “check assumptions” on one task later begin to check assumptions spontaneously in another domain, the support has gone beyond temporary improvement in task performance. Conversely, if they become careful only when AI is present and return to their previous style of judgment when support disappears, the system may be functioning more as an external aid than as capability formation. Neither is inherently bad, but claims about the effect should distinguish between them.
Decision time also cannot be ignored. Support for metacognition often requires additional checking. If accuracy improves only slightly while time increases substantially on simple, low-risk tasks, its practical value may be limited. In decisions where serious losses are possible, by contrast, additional time may be acceptable. Effects therefore need to be evaluated not only by accuracy, but in combination with task risk, time, and cognitive load.
We must also examine the possibility that support increases overconfidence. The fact that “AI even checked opposing views” may give users excessive reassurance. Even when a system presents multiple perspectives, the underlying information may be wrong or important perspectives may still be missing. If the existence of metacognitive support itself is taken as a guarantee of quality, the support function becomes a new authority.
Conversely, excessive warnings may make judgment impossible. If AI presents numerous counterexamples, uncertainties, and additional conditions every time, users may continually feel that “we still have not examined enough.” Research therefore needs to vary not only the presence or absence of support, but also the frequency and timing of intervention. A method that intervenes only when there are signs of error may have a different relationship between accuracy and burden from a method that intervenes constantly.
Comparison conditions are also important in experimental design. Comparing AI support with no support alone makes it difficult to separate the effect of simply receiving more information from the effect of metacognitive support. Conditions that provide the same amount of additional information, provide only opposing views, or require participants to list assumptions themselves can help examine in greater detail which intervention affects what. Effects may also differ according to task type and participant experience.
We also need to distinguish short-term judgment performance from long-term capability formation. If AI compensates for a weakness in the moment, immediate outcomes may improve. But whether long-term use increases users’ own ability to examine judgments or instead increases dependence on external support is a different question. This difference can be studied by evaluating users without support after sustained use and by measuring transfer to different tasks.
When evaluating a design such as Sai＊aus that emphasizes updates in understanding, self-report that “the user felt they thought more deeply” is not sufficient. We need to measure under what conditions error detection improved, whether confidence calibration changed, whether revision of assumptions transferred to other tasks, and whether agency in judgment was preserved. At the same time, it is important not to assume improvement in every situation, but to record conditions in which effects are weak or adverse.
The value of AI that supports metacognition is not established by an attractive principle alone. Only by testing when, on which tasks, and according to which measures the ability to reconsider one’s own thinking produces improvement can design claims be connected to empirical effects. There is a distance that must be measured between AI appearing to “make people think” and human judgment actually improving. Researching while preserving that distance is the caution required of a design that claims to support understanding.
Not answers. Understanding.