Analogy is used to transfer relationships from a known domain to another domain and create new hypotheses or design proposals. Because AI can generate candidates from broad knowledge, it can quickly present cross-domain similarities that humans may find difficult to conceive. Yet the statement “A is like B” alone does not tell us what is similar, which differences were ignored, or how far the similarity can be used. To make creativity examinable, the abstraction operation that produced the similarity needs to be decomposed.
In abstraction, some of the many attributes of an object are set aside while particular relationships are preserved. When comparing a road network with a communication network, for example, we can abstract away material and physical size while preserving the relationship “connecting multiple points through routes and distributing flows.” Once this operation is made explicit, the basis of the similarity changes from “they somehow resemble each other” to “they share this relational structure.” At the same time, we can examine whether the attributes that were abstracted away are truly irrelevant to the problem.
The first step toward auditability is to make correspondences explicit. We show what elements in the source domain correspond to in the target domain, which relationships are preserved, and which are not. Next, we separate predictions derived from the analogy. The presence of a shared structure does not mean that every property holding in the source domain transfers to the target domain. Distinguishing structural correspondence from additional inferences drawn from it makes the location of the leap visible.
Failure conditions are especially important. If we can show under what conditions an analogy remains valid and where it ceases to work, it becomes easier to distinguish a creative metaphor from a mistaken identification. When an ecosystem is compared to an organization, relationships such as interdependence and diversity may be useful, but mechanisms of biological selection cannot simply be transferred into norms for human organizations. The value of a similarity depends not only on the number of similar points, but also on the clarity of the boundary between relationships that can and cannot be transferred.
If we ask AI to output this structure, it becomes easier to examine when the source structure, target correspondence, preserved relationships, abstracted-away attributes, failure conditions, and hypotheses derived from the analogy are separated, rather than merely asking it to “give a similar example.” This does not directly visualize AI’s internal computation. It is a method for transforming a generated analogy into an external representation that humans can examine. Preserving this limitation is important in order not to overstate explainability.
Making the abstraction operation visible also changes how creativity is evaluated. We can ask not only whether a combination is unusual, but what structural commonality the combination rests on and whether it generates new predictions or design judgments in the target domain. Conversely, an analogy with high surprise value but vague correspondences and no identifiable failure conditions may stimulate ideation without constituting strong grounds for inference.
This method does not mean fully formalizing human ideation. Creative insights do not always appear from the outset as clear procedures. What matters is being able, after an insight has emerged, to examine how far the similarity can be trusted and where it can be applied. By preserving freedom at the generation stage and externalizing the content of the abstraction at the evaluation stage, creativity and examinability can be handled without placing them in opposition.
In human-AI collaboration, one possible division of labor is for AI to generate candidate similarities from a broad search space while humans evaluate their relevance to the purpose and their ethical implications. Yet here too, roles need not be fixed. AI can search for failure conditions in a similarity proposed by a human, and a human can revise correspondences proposed by AI. What matters is sharing not only the final metaphor, but also the abstraction through which that metaphor became possible.
The power of analogy does not lie in making distant things appear identical. It lies in temporarily extracting one relationship as a common structure while preserving the differences. If that operation can be made visible, the large number of similarities generated by AI can be transformed from mere idea candidates into intellectual resources that can be examined. The next challenge for creativity is not the quantity of combinations, but the ability to explain what was treated as the same and to examine the limits of that judgment.
Not answers. Understanding.