Must AI Only Produce Answers? Rethinking AI Through Zen Koans

🎥 The YouTube version is also available:

Beyond Answers: What Zen Koans Teach Us About AI and Deeper Thinking

Books: INTELLIGENCE FIELD: Intelligence as Relationship

When using generative AI, we tend to seek “more accurate answers.” An AI that responds quickly and plausibly to a question, with supporting evidence, is certainly important.

But in corporate decision-making, research, education, and creative work, is the ability to produce answers all that is truly needed?

Zen kōans offer a valuable perspective for thinking about this question.

A kōan is not a problem with one correct answer

Zen dialogues, or kōans, are not knowledge questions designed to identify a single correct answer.

Take the kōan, “Listen to the sound of one hand clapping.” There is no model answer in the ordinary sense. The more one tries to build up logic, construct an explanation, and fix the conclusion in words, the more one reaches an impasse.

That impasse itself has meaning.

What is called into question is the stance of the self that searches for an answer as an external object, divides the world through language, and seeks to possess the correct conclusion. The value of a kōan lies less in the answer itself than in the process through which the assumptions underlying one’s approach to an answer are unsettled.

Of course, Zen is not concerned with introspection alone. It also emphasizes how direct insight prior to conceptualization can transform everyday action, relationships with others, and the way one makes judgments.

This structure offers important implications for AI design as well.

Today’s AI is optimized as a machine for producing answers

Generative AI today excels at returning the most plausible output in response to a given prompt. In search, summarization, classification, writing, and code generation, this capability creates substantial value.

But the difficult cases in real work are not limited to situations where the question is clearly defined.

• The framing of the problem itself may be wrong.
• The underlying data or classifications may be outdated.
• The interests and values that should take priority may not yet be clear.
• The basis for a decision and the acceptance of stakeholders may matter more than the correctness of the answer itself.

In such situations, an AI that merely generates fluent answers can rapidly amplify false assumptions. The more powerful AI becomes as a “correct-answer generator,” the easier it becomes for people to accept its output without questioning it.

What is needed is an expansion from AI that produces answers to AI that can also work with the conditions of a question and the process of judgment.

Proposal: From AI as an answerer to AI as an interactive space for updating assumptions

There is no need to transplant Zen kōans directly into AI. But the function of a kōan—bringing assumptions into view and prompting their reconsideration—can become a design principle for AI systems.

1. Present assumptions alongside the answer

AI should present not only its conclusion, but also the information, assumptions, and evaluation criteria on which that conclusion depends.

For example, when recommending a management initiative, the AI should indicate whether it prioritized revenue growth, customer satisfaction, or the acceptable level of operational burden on frontline staff. Users can then do more than agree or disagree with the conclusion; they can revise the conditions on which the judgment was made.

The point is not to perform a plausible-looking “chain of thought.” It is to connect the judgment to verifiable evidence, the scope of available data, permissions, and the model’s limitations.

2. Give AI a responsibility to ask back

When a question is ambiguous, contradictory, or high-risk, AI should not simply fill in the gaps and produce an answer. It should return the issues that require confirmation.

Even in response to a question such as, “How can we increase sales?”, an AI could ask: “Which should take priority: short-term revenue, retention, or customer trust?” The purpose of such questions is not to prolong the conversation. It is a safety mechanism for making overlooked choices visible and preventing automation from advancing on the basis of mistaken assumptions.

3. Make judgments updateable rather than ending them with a single output

Real-world decision-making does not end when a conclusion is reached once. New information, responses from the field, and unforeseen side effects can all require the decision to be updated.

AI needs mechanisms not only to provide an initial answer, but also to trace which judgments changed, when they changed, and on what evidence. Records such as a Decision Trace, Evidence Graph, and Retrieval Receipt are not merely audit logs. They become an organization’s memory for relearning from its own decisions.

4. Design relationships rather than relying on a single intelligent AI

In complex organizations, it is risky for AI to authoritatively present one final answer. This is because the people who hold knowledge, those who hold decision rights, those affected by the decision, and those who audit it each have different perspectives and responsibilities.

The role of AI, therefore, is not to become an all-powerful arbiter. It is to connect multiple perspectives, data sources, and forms of expertise; to avoid concealing conflict and uncertainty; and to make the boundaries of judgment clear.

What matters here is not model performance alone. It is also institutional design: permission management, boundaries around accessible knowledge, preservation of evidence, approval processes, mechanisms for objection, and the ability to withdraw or reverse decisions.

An AI that does not answer immediately is not a weak AI

An AI that does not provide an immediate answer is not necessarily incapable. On the contrary, the ability to distinguish between questions that should be answered and questions whose assumptions should first be examined is part of advanced intelligence.

Of course, not every situation requires a philosophical dialogue. For routine work and clearly defined inquiries, an AI that answers quickly and accurately provides the greatest value.

But for decisions with significant consequences, decisions involving conflicts of values, and decisions based on incomplete data, AI should work toward improving the quality of judgment rather than merely automating conclusions.

Zen kōans do not tell us to discard answers. Rather, they can be read as urging us not to lose sight of the stance from which we arrive at an answer.

The same is true for AI. The goal should not be an intelligence that calculates a single correct answer and stops there. It should be an intelligence that enables people and organizations to reexamine their own assumptions, boundaries, relationships, and decision processes—and to update them when necessary.

Such AI can evolve from a machine that produces answers into a social intelligence infrastructure that cultivates better questions and better relationships.

Chinoba

Chinoba

Intelligence as Relationship

Research Platform

founded by

Masao Watanabe

AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance