“Explanatory Lock-In” in Technology

Why the First Widely Accepted Explanation Often Survives Even After the Technology Itself Has Changed

When a new technology emerges, society needs a simple explanation in order to understand it. A self-organizing map (SOM) may be described as “a learning method that gradually updates the node closest to the input data and its neighbors.” RAG may be explained as “a mechanism that retrieves external documents and provides them to an LLM.” An LLM may be described as “a model that predicts the next token.” None of these explanations is entirely wrong. In fact, they are highly effective for conveying the essential features of a new technology in a short period of time, which is precisely why they spread so widely.

The problem arises when the technology itself evolves, its intended uses change, and its implementation methods improve, while the original explanation continues to survive as though it still represented the essence of the technology.

In this article, I will refer to this phenomenon as “Explanatory Lock-In.” This is not an established term in psychology, but the mechanism can be understood reasonably well by combining several known concepts: the primacy effect and continued influence effect in psychology, information cascades in the social sciences, and path dependence in the economics of technology. Once we also take into account textbooks, education, software implementations, search engines, and, more recently, the reproduction of information by LLMs, explanatory lock-in should be viewed as more than an individual cognitive bias. It is better understood as a form of social self-reproduction of knowledge.

Why Is the First Mental Model So Difficult to Revise?

The primacy effect refers to the tendency for information presented first to exert a disproportionately strong influence on later judgments and impressions. Human beings do not necessarily store each piece of information as an isolated fragment. Instead, we tend to form an initial conceptual model of an object and then interpret later information through that model. The first explanation therefore functions not merely as the first item of information received, but as a framework that organizes subsequent understanding.

A related phenomenon is the continued influence effect. This describes the tendency for misinformation that has once been accepted to continue affecting reasoning and judgment even after it has been corrected. The important point is not simply that people fail to forget outdated information. Rather, once they have constructed a causal model or narrative based on that information, removing one element leaves a gap. To abandon the old explanation, they must construct an alternative explanatory framework.

Technology presents an even more difficult case, because the original explanation is often not false. It may have been perfectly reasonable when the technology first appeared, only to become incomplete as the technology later evolved.

The challenge, therefore, is not simply to replace “wrong information” with “correct information.” It is to replace a correct but limited model with a broader and more appropriate one. That can be considerably more difficult than correcting a straightforward error.

Explanations Become Embedded in Social Systems, Not Just in Individual Minds

Explanatory lock-in cannot be understood through psychology alone. Explanations of technology become embedded in textbooks, lecture materials, technical blogs, sample code, open-source libraries, APIs, product manuals, and search results. Once an explanation becomes the standard account of a technology, it is taught in education, encoded into software, and reproduced by users of that software.

In the economics of technology, path dependence describes how early choices can constrain later choices. Once a technology gains widespread adoption, user growth, complementary products, accumulated expertise, and standardization make it increasingly rational to continue using it, even if superior alternatives later emerge. A similar phenomenon can occur not only with the technology itself but with the explanation of the technology.

A simple explanation spreads first. Textbooks adopt it. Developers who learned from those textbooks create software based on that understanding. Users of the software internalize the same explanation. Over time, the explanation becomes not merely one interpretation among many, but part of the shared assumptions of the technical community.

There is also the problem of information cascades. If one hundred sources repeat the same explanation, it may appear to have overwhelming support. But those one hundred sources may not represent one hundred independent judgments. They may all derive from the same textbook, the same foundational paper, or the same official tutorial.

The fact that an explanation is widely repeated does not mean it has been independently verified many times.

Explanatory lock-in can therefore be understood as a composite phenomenon involving the primacy effect, continued influence, information cascades, and path dependence.

SOM as an Example of Explanatory Lock-In

The Self-Organizing Map (SOM) provides a useful example.

SOM became widely known in the early 1980s as a sequential learning algorithm in which data points are presented one at a time. For each input, the closest node is identified, and both that node and its neighbors are incrementally updated. This mechanism is easy to explain. An input stimulus arrives, the closest neuron becomes the winner, neighboring neurons move with it, and repetition gradually produces an organized map. The explanation also fits naturally with the image of neural organization in the brain, making it an intuitively powerful story.

Later, however, batch-learning algorithms for SOM were introduced. For statistical data analysis, batch learning is often preferable because it can reduce dependence on data presentation order and learning-rate schedules. Yet the explanation “SOM means presenting data one by one and gradually moving the winner and its neighbors” remained dominant for many years. Among researchers, software developers, and users, the sequential learning procedure itself was often treated as though it were the definition of SOM.

The important point is that the original explanation was not wrong. It correctly described one implementation of SOM. What happened was a confusion between the higher-level objective of the method and the specific algorithm used to achieve it.

If the purpose of SOM is understood more broadly as arranging representative vectors while approximately preserving topological relationships in the data space, then sequential learning and batch learning are simply different means of achieving that objective. But once people learn the technology primarily as a procedural recipe, the procedure itself easily comes to be mistaken for the essence of the technology.

This problem is reinforced by education. Textbooks must explain procedures, and in doing so they often remove the context in which the technology was developed, the problem it was intended to solve, and the reasoning behind the design. Students then learn how to execute the method rather than why the method exists.

The means gradually becomes the purpose.

In RAG, “Retrieval” Risks Becoming the Goal

A similar pattern can be seen in Retrieval-Augmented Generation (RAG).

The underlying idea of RAG is to supplement the knowledge inside a generative model by referring to external information when internal knowledge is insufficient. As RAG became widely adopted in practice, however, it increasingly came to mean a specific implementation pattern:

documents are divided into chunks, embeddings are computed, the chunks are stored in a vector database, the most similar chunks are retrieved using top-k search, and those chunks are inserted into the prompt given to the LLM.

This is a perfectly legitimate implementation of RAG. But in enterprise AI, the information required for a sound decision is not always “the text most semantically similar to the question.”

A system may need to determine which policy is currently valid, what history exists with a particular customer, what the current state of a project is, which document is authoritative, whether an exception applies to a general rule, or how conflicting information should be resolved.

These are not merely retrieval problems. They are closer to problems of context reconstruction: rebuilding the relevant state of the world at the moment a decision must be made.

If people become locked into the explanation “RAG equals vector search,” then when problems arise, improvement efforts tend to remain inside the same framework: change the embedding model, adjust chunk size, add a reranker, increase the number of retrieved documents.

But the more fundamental question may be whether the problem should have been formulated as a retrieval problem in the first place.

This is where explanatory lock-in moves beyond a problem of understanding and begins to affect technology investment and system design. The organization may not simply be choosing the wrong method. It may be defining the problem incorrectly because an outdated explanation has framed the problem in the wrong way.

The LLM Explanation: “It Just Predicts the Next Token”

The same issue is emerging around LLMs.

One of the most common explanations of modern large language models is that they are “models that predict the next token.” This captures an important feature of their training process and is not incorrect. But once again, there is a danger of conflating the training objective with the internal representations and capabilities that emerge after training.

From the fact that an LLM is trained through next-token prediction, it does not automatically follow that it “does not deal with meaning,” “does not reason,” or is merely a “statistical parrot.” Such conclusions infer the total nature of the system directly from its training objective.

The opposite exaggeration is equally problematic. The fact that a model produces fluent language does not prove that it possesses the same forms of understanding, cognition, or consciousness as a human being.

What is needed is not to lock the LLM into a single explanation.

From one perspective, it is an autoregressive probabilistic language model. From another, it is a large-scale representation-learning system. From a systems-engineering perspective, it can also be treated as one component in a broader architecture involving retrieval, reasoning, tool use, memory, and external state.

Each of these explanations captures one aspect of the system. None alone necessarily defines the whole.

If the statement “an LLM is ultimately just a next-token predictor” remains the dominant explanation for decades, we may eventually see the same phenomenon that occurred with SOM: the technology itself will have changed substantially while the social understanding of it remains frozen at an early stage.

Education May Produce “Degraded Copies” of Explanations

Education is particularly important in understanding explanatory lock-in.

A textbook is an extreme compression of knowledge that researchers may have spent decades discovering, debating, testing, and refining. During that compression process, much of the original context disappears: what was once unknown, why the problem mattered, which hypotheses failed, where the theory breaks down, and what kinds of uncertainty remain.

The knowledge may remain technically correct while losing the process that produced it.

When the original researcher, or someone closely connected to the research, explains a phenomenon, they can often begin with the underlying question: What was strange? What did not make sense? Why was a new concept needed?

When the knowledge is taught several steps removed from its source, the explanation is more likely to become a sequence of definitions, formulas, procedures, and exercises.

The original transmission path of knowledge is:

phenomenon → recognition of a problem → trial and error → hypothesis → theory → textbook

But within an educational system, it often becomes:

textbook → lecture materials → classroom explanation → examination

The definitions and formulas may remain intact, yet the motivation, limitations, tacit knowledge, and historical problem-setting gradually disappear.

If we call this a “degraded copy,” what degrades is not necessarily the factual accuracy of the content. What degrades is its meaning.

From this perspective, the quality difference among educational institutions cannot be reduced simply to differences in curriculum. One important factor is the distance from the place where knowledge is actually being created.

Learning from someone actively conducting frontier research, learning from someone who understands the textbook well, and learning from someone who teaches a summary of the textbook may all cover “the same subject,” yet the resolution of the knowledge can be very different.

The closer the teaching is to the site of knowledge creation, the more likely students are to hear statements such as:

“Textbooks usually explain it this way, but the real problem is here.”

That kind of knowledge rarely survives intact in finished textbooks.

LLMs May Strengthen Explanatory Lock-In

Until recently, outdated explanations were reproduced mainly through textbooks, lectures, and technical documentation. LLMs now make that reproduction dramatically faster.

A dominant explanation on the web becomes part of the data used to train LLMs. Users then ask the LLM to explain the technology, and the model reproduces the dominant explanation. Those answers are then turned into blog posts, internal documents, presentation materials, and new web content.

A feedback loop can emerge:

existing explanation → LLM training → LLM output → new articles → future training data

The problem is that this loop does not guarantee the updating of knowledge.

LLMs are extremely good at reconstructing explanations that are already widely shared. But unless explicitly required to investigate historical changes, later revisions, conflicting interpretations, and primary sources, they may simply reinforce the majority account.

In the AI era, therefore, statements such as “search results say this,” “the LLM answered this way,” or “many articles say the same thing” are not sufficient evidence that an explanation remains valid.

What matters is when the explanation was formed, which primary sources it ultimately derives from, and how the technology has changed since then.

Effective Use of Technology Requires Escaping Explanatory Lock-In

This issue is not merely of historical or academic interest.

When companies adopt a new technology, they reasonably ask questions such as, “What is the standard approach?” or “What are the best practices?” But in rapidly evolving fields, the “standard approach” may not mean the method that is currently most appropriate. It may simply mean the method that became most widely adopted earlier.

The danger becomes particularly clear when a technology fails to produce the expected results.

If an organization is locked into the existing explanation, it will tend to search for improvements within the same conceptual framework. If RAG performs poorly, the response is to improve retrieval. If an SOM produces unstable results, the response is to tune the learning rate.

Yet the real question may be whether the problem should be treated as retrieval in the first place, or whether sequential learning should have been assumed at all.

To use technology effectively, organizations must therefore examine not only the algorithm but also the conceptual model that led them to choose that algorithm.

In other words, keeping up with technological change requires not only the ability to update technology, but also the ability to update explanations of technology.

How to Escape Explanatory Lock-In

The first step is to distinguish between the definition of a technology and a representative implementation. RAG is not identical to top-k retrieval from a vector database. SOM is not identical to a sequential learning algorithm. Asking, “Is this the technology itself, or merely one implementation?” can immediately broaden the conceptual space.

Second, technology should be examined historically rather than only at the moment of invention. Reading the original paper is important, but not sufficient. It is useful to distinguish at least three stages: the technology at invention, the technology at mass adoption, and the technology today. Later review papers, revisions by the original authors, and operational failures may reveal important changes that never became part of the popular explanation.

Third, the genealogy of explanations should be examined. If the same explanation appears in one hundred sources, those sources may all descend from the same paper, textbook, or tutorial. The number of sources is less important than the number of genuinely independent intellectual lineages behind them.

Fourth, failed cases should be studied carefully. Successes tend to reinforce the dominant explanation, while failures reveal its boundaries. An unsuccessful implementation may show that the problem lies not in the technology itself but in the conceptual model used to frame the problem.

Fifth, the problem should be redefined one level higher. Instead of asking, “How can we improve retrieval accuracy?” ask, “What information state must the AI reconstruct in order to make a sound decision?” Instead of asking, “What learning rate should SOM use?” ask, “What learning method best produces a stable structural representation?” Raising the level of abstraction makes it easier to escape implementation-specific assumptions.

Sixth, explanations should be treated as having versions. Software has version numbers, but conceptual explanations are often treated as permanent definitions. There is no reason why “RAG as understood in 2020” and “RAG as understood in enterprise AI in 2026” should be identical. Explanations should themselves be treated as models that require revision.

Seventh, multiple explanations should be held simultaneously. An LLM can be viewed as a probabilistic language model, a representation-learning system, a reasoning component, or a language interface within a larger enterprise system. Different levels of explanation are useful for different purposes.

The goal is not to find one perfect description that permanently replaces all others, but to avoid treating any single explanation as an exclusive and final truth.

Technological Innovation Also Requires Innovation in Explanation

We tend to seek “the correct explanation” of a technology. But in rapidly evolving fields, it may be more important to maintain the most useful explanation available today while remaining prepared to revise it tomorrow.

Simplification is unavoidable. Early explanations are necessary. The problem is not simplification itself. The problem begins when we forget that we simplified.

Sequential learning in SOM, vector retrieval in RAG, and next-token prediction in LLMs each capture important aspects of the technologies they describe. But the moment any of these becomes the final answer to the question “What is this technology really?”, understanding begins to fall behind the technology itself.

A widely shared explanation of a technology may not describe the technology as it exists today. It may simply preserve the conceptual model that was formed when the technology first entered public awareness.

Mature technological understanding therefore requires more than knowing what most people say. It requires asking:

When was this explanation formed? What problem was it originally intended to solve? How has the technology changed since then? And is this still the best way to explain it today?

In an era when AI accelerates technological change while simultaneously reproducing existing knowledge at enormous scale, this ability becomes increasingly important.

The challenge ahead is not merely to follow new technologies. It is to identify explanations that have become obsolete, discard them when necessary, and sometimes redefine the problem itself.

Technological innovation is not only innovation in devices and algorithms. The conceptual models through which we understand technology must also continue to evolve.