For AI to handle large numbers of concrete examples, it is not enough to process every input as a completely independent case. If features and relationships shared across different inputs can be extracted and used as higher-level representations, existing structures can be transferred even to situations not previously encountered. In this sense, abstraction is not merely information reduction; it is an important function that enables reasoning and generalization. But when human experience is the object, another problem arises. Even if an experience can appropriately be summarized under one higher-level concept, that does not mean it is the only abstraction that holds for that experience.
For example, the statement “I want to leave my current job, but I am afraid of disappointing those around me” permits multiple abstractions. It can be represented as a relation between self-determination and responsibility, as a relation between recognition by others and self-evaluation, or as the effect of a group’s norms on an individual’s choice. It can also be understood as avoidance of uncertainty about the future. Because all of these are derived from the same statement, they may look redundant, but they are not synonymous. Each brings different relationships to the foreground, treats different conditions as important, and affects subsequent questions and judgments differently.
The issue, then, is whether we should simply select the one abstraction with the greatest explanatory power. Integration into one representation simplifies processing and makes mapping to similar cases easier. But if everything is integrated under “recognition,” economic uncertainty or changes in belonging may recede from subsequent reasoning. Conversely, if “uncertainty” is placed at the center, the relation by which the person has internalized others’ expectations may appear weaker. Efficiency through abstraction must be evaluated not only by what was retained, but also by what was removed from the space of exploration.
Accordingly, when considering internal efficient representation for human experience, the direction of “compressing many concrete examples into a small number of concepts” is not sufficient. When multiple nonequivalent higher-level representations hold for the same experience, they need to be retained in relation to one another without being prematurely integrated into a single superordinate concept. Coordinates of freedom, recognition, and norms need not be mutually exclusive. At the same time, combining them into an overly broad concept such as “a human-relations problem” would erase the structures each representation distinguished. Efficiency is not synonymous with integration that eliminates difference.
To think about this structure, we must distinguish “producing multiple answers” from “retaining multiple internal representations.” It is possible to generate many superficially different sentences from one internal classification. In that case, the outputs are diverse while the way the problem is represented remains singular. Retaining multiple higher-level representations, by contrast, means retaining states that differ in which conditions they treat as important, what they place at the causal or relational center, and what additional information would update the current interpretation. What matters is not the number of representations, but their nonequivalent roles in reasoning.
Philosophical concepts provide strong material for considering this nonequivalence. Across the history of philosophy, objects such as freedom, desire, time, others, norms, and finitude have not converged on single definitions; different relational structures have been proposed from different premises. These disagreements need not be treated as disorganized knowledge or contradictions to be eliminated. If they are used as resources showing that the same human experience can take different structures in multiple spaces of meaning, philosophy does not add correct answers to AI; it provides comparable coordinates for examining premature convergence on a single representation.
However, merely using the phrase “philosophical coordinates” does not establish that corresponding conceptual structures exist inside AI. A model’s ability to generate words such as “freedom,” “recognition,” and “norms” appropriately is separate from whether those representations play different causal roles in subsequent reasoning. To claim them as internal representations, tests are needed: which representations recur stably when input conditions change, whether nearby representations can be distinguished, and whether intervention on a particular representation changes tendencies in judgment or generation. Naming a concept may begin an analysis, but it does not prove an internal mechanism.
Retaining multiple higher-level representations also has a cost. If candidates continue to multiply, the search space expands and reasoning becomes less efficient. The design problem is therefore not to preserve unlimited possible interpretations, but to determine under what conditions nonequivalent representations that could affect the current judgment should be retained and under what conditions they can recede into the background. Even when a representation is backgrounded, it need not be deleted if it can be used again when new information arrives. Here, re-expandability and updateability become part of efficiency alongside compression ratio.
Defined from this perspective, the goal of internal efficient representation changes from “covering the maximum number of cases with the minimum representation.” More important is to suppress complexity unnecessary to current reasoning without losing differences in meaning that could alter future judgment. If common parts can be shared while only nonequivalent relationships are retained as branches, every case need not be stored independently. At the same time, there is no need to integrate irreversibly into a single abstract concept. Efficiency can also be evaluated by how little burden is required to maintain a structure that can return to necessary differences.
Applied to the design of Sai＊aus, the purpose of retaining multiple philosophical coordinates is not to generate as many answers as there are philosophers. It is to distinguish how the problem setting changes depending on which relationship is foregrounded for the same input, and connect those differences to a form in which humans can judge them. If self-determination is central under one coordinate and social norms under another, AI should not determine either as the person’s “true problem,” but handle which conditions each explains and which it does not. Multiple internal representations are used not to show humans more information, but to prevent premature fixation of meaning.
This design must also be positioned carefully in comparison with general-purpose AI. Large-scale AI can generate multiple perspectives and present different explanations according to context. The difference therefore cannot be stated simply as “other AI gives only one answer.” The design focus Sai＊aus can place is not the mere ability to generate multiple views, but treating them as nonequivalent problem representations and explicitly designing which differences to retain, under what conditions to foreground them, and where to return meaning judgment to the human. What should be compared is not the number of perspectives, but how differences among perspectives connect to updates in understanding.
Evaluation methods must change to match this design. Candidate measures include whether applying different coordinates to the same input actually changes the selection of important conditions, whether the weighting of appropriate coordinates can be updated with additional information, whether a backgrounded representation can be re-expanded when needed, and whether presenting multiple representations enables users to distinguish major decision points more accurately. These are not one universal scale, but they can connect the design principle of “multiple representations” to observable behavior. It is important to place a measurable distance between principle and effect.
A single abstraction is often powerful. Without finding common structure, knowledge transfer and efficient reasoning become difficult. Yet in human experience, the very success of an abstraction can erase another meaning. What is needed, therefore, is neither rejecting compression nor preserving complexity unchanged. It is handling how far things can be shared and where they should branch into nonequivalent representations. Multiple philosophical coordinates offer one way to think about that boundary.
For AI that seeks to understand humans, efficiency is not only reaching one explanation by the shortest route. It also includes foregrounding the structure needed for the current judgment without losing the possibility of future updating through another perspective. Instead of a single minimal representation, maintain representations that can re-expand into different spaces of meaning when necessary. Internal efficient representation then changes from compression that eliminates complexity into a design that preserves important differences within complexity in reusable form. And which meaning is ultimately taken up as one’s own problem is not fixed inside AI, but returned to human judgment.
Not answers. Understanding.