Human Intuition, Collective Illusion, and Concept-Structure Analysis in the Age of AI
I have long been somewhat skeptical of the KJ Method.
This is not to say that the KJ Method was entirely meaningless. The procedure of converting fragmented information into cards, grouping similar items, giving labels to each group, and diagramming the relationships among them certainly had some practical value in organizing chaotic information. In particular, within organizations, it was useful for allowing members to share what they felt, what they regarded as problems, and how they understood a given situation.
However, I have always felt a strong sense of discomfort with the way the KJ Method has often been described as a “method of creative thinking” or a “technique for collective intellectual production.”
To put it bluntly, the KJ Method may have been an adult version of the Ouija Board (Kokkuri-san in Japanese).
The Ouija Board and the KJ Method
In the 1970s, while companies were enthusiastically adopting the KJ Method, children were playing Kokkuri-san after school in Japan, a divination game similar to the Ouija Board. They would ask innocent or mischievous questions such as, “Who does so-and-so like?” or “What will happen in the future?” and watch as a coin or pencil moved, as if some external will were at work.
Of course, in reality, it was the participants themselves who were moving the coin. Their subtle unconscious movements, expectations, tension, and the atmosphere of the group appeared as if they were the will of a spirit.
The KJ Method contains a similar danger.
Participants arrange cards, gather similar ones, give names to groups, and diagram the whole. In this process, their intuition, assumptions, expectations, experiences, power relationships, and the atmosphere of the room play a major role. Yet once the process is completed, the resulting structure may appear as if it had “naturally emerged from the data itself.”
A structure created by humans is mistaken for an objective structure discovered by humans.
This is the fundamental danger of the KJ Method.
The KJ Method Does Not Produce Genius
The KJ Method is a way of externalizing and relating scattered pieces of information. In this sense, it can be seen as an attempt to bring onto paper the classification and association that humans unconsciously perform in their minds.
The critic Tachibana Takashi, as I recall, criticized the KJ Method along these lines: it externalizes what everyone is already doing inside their heads, but does so in a manner as inefficient as running a three-legged race. My own view is close to this.
Inside the human mind, classification, association, rearrangement, and abstraction occur at high speed. Writing these processes out on cards, arranging them on a table, looking at them with several people, and forming “islands” out of them is also a procedure that converts individual intuitive thinking into a slow collective exercise.
Of course, this had meaning within organizations. The reason is simple: in an organization, what is inside each person’s head is not visible. Turning thoughts into cards and grouping them can help members share what they are thinking, what they are dissatisfied with, and where they feel something is wrong.
But this is not so much “intelligence amplification” as “recognition sharing.”
Even if ordinary people gather and practice the KJ Method, it does not mean that genius-level results will be produced. Ordinary observations produce ordinary cards. Ordinary cards produce ordinary clusters. Ordinary clusters receive ordinary labels.
A methodology is not magic that automatically allows practitioners to transcend their own intelligence.
GTA Does Not Promise Genius Either
The same can be said of the Grounded Theory Approach, or GTA.
GTA is stricter than the KJ Method. It includes procedures such as open coding, categorization, constant comparison, theoretical sampling, memo writing, and saturation. As a qualitative research methodology, it has been institutionalized to a certain degree.
If the KJ Method is a method that permits ambiguity and human intuition, GTA is a method that imposes more rigorous rules on the researcher.
However, rigorous rules do not necessarily expand creativity. In fact, they may have the opposite effect. The more faithfully researchers try to follow the procedure, the more they may become bound by “correct coding,” “valid categories,” and “explainable relationships.” As a result, leaps of thought, discomfort, heterogeneous associations, and undefined structures may become harder to notice.
GTA may increase the explainability of research. But it does not guarantee genius-level insight.
Just as the KJ Method does not turn a group into a genius, GTA does not turn a researcher into a genius.
Yet the Two Methods Share an Important Common Structure
Does this mean that the KJ Method and GTA were meaningless?
No.
What the two methods share is that both attempt to divide scattered information into units, group them based on similarity, extract common characteristics from each group, and explain relationships among the elements.
In the KJ Method, this is done through cards, islands, labels, and diagrams.
In GTA, it is done through codes, categories, concepts, and theory.
Their surface forms are different. But the underlying process is quite similar.
- Extract fragments of information.
- Observe similarities among them.
- Group similar items together.
- Name the groups.
- Consider the relationships among the groups.
- Explain the overall structure.
This is a form of semantic clustering performed manually by humans.
And this was the first intuition behind my approach to concept research.
The Origin of Concept Research
What the KJ Method and GTA perform through human intuition and procedure may, in essence, be a machine-learning-like process.
Treat scattered information as data points.
Group them based on similarity.
Extract the common characteristics of each group.
Represent the relationships among groups as a structure.
Interpret that structure, explain it, and connect it to decision-making.
This corresponds to clustering, feature extraction, profiling, dimensionality reduction, network analysis, and model interpretation.
The KJ Method and GTA can be understood as forms of “unsupervised learning of meaning” that humans have carried out intuitively and experientially.
This intuition gave rise to the idea of concept research.
Could the grouping and structuring of meaning that humans have performed vaguely and intuitively be externalized through machine learning? Furthermore, if that structure could be made readable by humans and referable by AI, could we create a shared conceptual space between humans and AI?
This question became the origin of GNG+MST Concept-Structure Analysis, and later ConceptMiner.
GNG+MST Is Not an AI Version of the KJ Method
The important point here is that GNG+MST Concept-Structure Analysis is not simply an AI-enhanced version of the KJ Method.
In the KJ Method, humans gather cards that they somehow feel are “close.”
In GTA, researchers decide that “this belongs to this category.”
Both ultimately depend on human judgment.
In GNG+MST Concept-Structure Analysis, texts, documents, and concepts are represented as embedding vectors, and their semantic proximity is treated as a spatial structure. GNG, or Growing Neural Gas, extracts the distributional structure within that space. MST, or Minimum Spanning Tree, reveals the connectivity among concepts.
Of course, this is not absolute truth. The semantic space created by an embedding model is not the world itself. It depends on the model, the data, preprocessing, and the distance metric.
Nevertheless, unlike the KJ Method or GTA, it is at least possible to state: “With this model, this data, and this distance structure, this conceptual arrangement appears.” This creates the possibility of critique.
The KJ Method externalized human intuition.
GTA proceduralized the researcher’s interpretive process.
GNG+MST externalizes the structure of semantic space in a computable form.
This is a major difference.
Avoiding the AI-Era Version of the Ouija Board
Here we must be careful: AI can also become a form of Ouija Board.
When an LLM fluently classifies, summarizes, labels, and offers strategic implications, humans can easily believe it. But is there really a structure there? Or is the LLM merely producing plausible language? The distinction is difficult.
In the KJ Method, humans mistook the structure they themselves created for a structure that had naturally emerged from the data.
In the age of AI, there is a risk that we will mistake interpretations generated by LLMs for objective structures discovered by AI.
If that happens, the human version of the Ouija Board will merely be replaced by an AI version of the Ouija Board.
That would be meaningless.
This is precisely why we should not rely solely on the linguistic interpretations produced by LLMs. We must place those interpretations alongside structures such as embedding spaces, distances, clusters, topologies, and MSTs, so that the LLM’s interpretations can be critically examined.
GNG+MST Concept-Structure Analysis is a method for doing exactly this.
A Semantic Interface Between Humans and AI
What I expect from GNG+MST is not merely support for recognition sharing among humans, as in the KJ Method.
What matters more is the sharing of semantic space between humans and AI.
AI processes text within high-dimensional semantic spaces. Humans cannot directly see those spaces. Even if AI judges that “this concept is close to that concept,” “this concept is peripheral,” or “this concept acts as a bridge,” the basis for that judgment usually remains a black box.
GNG+MST externalizes that semantic space as a conceptual structure that humans can understand.
At the same time, it can also express human intentions, values, and strategic judgments as a structure that AI can refer to.
In other words, GNG+MST Concept-Structure Analysis is not merely an analytical method.
It is a concept-sharing interface through which humans can understand the semantic space of AI, and AI can understand the intentional structure of humans.
Conclusion
The KJ Method was useful for recognition sharing within organizations. But it was not a method that turned groups into geniuses. Rather, it carried the danger of making structures created by the participants’ own intuition, assumptions, and group atmosphere appear as if they were objective structures naturally emerging from the data.
In that sense, the KJ Method was an adult version of the Ouija Board.
GTA is a stricter methodology than the KJ Method. But strictness does not guarantee creativity. GTA, too, does not promise genius-level results.
Nevertheless, the KJ Method and GTA share an important common structure. Both group fragmented information based on similarity, extract common characteristics, and attempt to explain relationships. This is a machine-learning-like process that humans have performed intuitively.
Concept research, and GNG+MST Concept-Structure Analysis, are attempts to reconstruct this human process of semantic structuring through machine learning, topology, graph structures, and LLMs.
The purpose is not to make the KJ Method more convenient through AI.
The purpose is to create conceptual structures through which humans and AI can understand each other’s semantic spaces and intentional structures.
If the KJ Method was a method for externalizing the intuition of human groups, GNG+MST Concept-Structure Analysis is a method for externalizing a semantic space that can be shared between humans and AI.