Technology, intuition and the future of work

Show original animation

How the Erosion of Tacit Knowledge Redefines the Boundaries Between Humans and Machines

Abstract: In a world where artificial intelligence advances at an accelerated pace, the boundary between tacit knowledge and explicit knowledge becomes the decisive terrain for understanding the future of work. Inspired by the ideas of Isaac Asimov, Michael Polanyi, and David Autor, we propose to explore how physical interaction, lived experience, and natural language processing (NLP) are reshaping human capabilities and the possibilities of automation. This analysis offers a conceptual framework to consider how technology may erode, but not entirely replace, the knowledge we cannot yet articulate.

In 1969, Isaac Asimov published a short story that, unintentionally, anticipated a dilemma now dominating the debate about artificial intelligence and employment: can a machine acquire tacit knowledge, that type of knowledge which is not formalized into explicit rules?

In Feminine Intuition, Asimov presents Jane, a robot designed not merely to process data, but to develop tacit knowledge through direct interaction with its environment. Jane is not programmed with closed instructions; she learns by observing, by living real-world situations at the Flagstaff astronomical observatory.

More than half a century later, science and economics revisit this question to understand the current limitations of technology. Michael Polanyi, almost contemporaneously with Asimov, stated his famous Polanyi's paradox: "we know more than we can tell." This insight is key to understanding why so many human tasks still resist automation, despite spectacular advances in AI and machine learning.

David Autor, one of today's leading labor economists, applies this idea to real-world dynamics: technology can easily automate tasks based on explicit rules, but struggles enormously with those relying on tacit skills, such as adaptability, contextual judgment, or common sense.

Inspired by these three thinkers, Asimov, Polanyi, and Autor, we developed a model aimed at capturing how physical interaction and direct experience with the world can erode the boundary between tacit and explicit knowledge, thereby altering the fate of millions of workers.

Physical Interaction and Real Learning

The model starts from a clear premise: there is a fundamental difference between processing structured data and learning by physically interacting with the environment. Jane, Asimov’s robot, needs to see, hear, and engage with real social contexts in order to develop tacit knowledge. Simply loading databases is not enough.

In the real world, even the most advanced AI systems stumble over the subtleties of human behavior and the unpredictable variations of complex physical environments. The transfer of tacit knowledge still requires direct interaction, tangible.

We model this process with a parameter J, which measures a system’s effectiveness at converting lived experience into codifiable knowledge. As J increases, the ability to transform tacit skills into explicit forms also grows.

In our simulation graphs, we observe how, as J rises, the possibility of automating tasks initially deemed unreachable for machines increases.

Tacit Knowledge vs. Explicit Knowledge

Polanyi’s paradox reminds us that not all knowledge is alike.

Explicit knowledge can be written down, taught step by step, codified into clear instructions. This is the type of knowledge that traditional automation can efficiently replicate: accounting rules, mathematical calculations, repetitive assembly.

Tacit knowledge, by contrast, resides in skills that are difficult to formalize: recognizing an emotion in a subtle gesture, improvising in an emergency, interpreting double meanings in conversation. Tacit knowledge is more often learned than taught, internalized rather than explained.

Our model initially measures how much of a task is tacit, using the parameter T. From there, it simulates how the effectiveness of J can gradually reduce the remaining tacit fraction and increase the explicit portion. The more a task’s knowledge can become explicit, the more likely it is to be automated.

This phenomenon is visualized in our three-dimensional surfaces, showing how the growth of J directly impacts the knowledge dynamics of tasks.

LLMs, NLP, and the Erosion of Tacit Knowledge

As language models (LLMs) and natural language processing (NLP) technologies evolve, their impact on tacit knowledge becomes increasingly profound. Trained on massive amounts of text, these systems are able to capture linguistic patterns, implicit contexts, and semantic structures that once depended solely on human experience.

LLMs do not merely imitate predictable answers; they begin to model aspects of implicit understanding. When a model can infer the meaning of an ambiguous statement, interpret irony, or suggest strategies based on incomplete descriptions, it is operating at the very edge of tacit knowledge.

However, this advancement does not occur spontaneously. It depends on two fundamental phenomena. First, corpus amplification: through the massive collection of human interactions—texts, dialogues, real-world scenarios—the model absorbs tacit forms of reasoning that, by becoming part of its training, are indirectly codified. Second, the emergence of generalized capabilities: LLMs do not simply memorize answers, but form latent representations that enable them to infer non-explicit connections.

Thus, the process of reducing tacit knowledge consists in transforming distributed, subtle experience into explicit structures, internalized within a statistical model. This phenomenon is analogous to our parameter J: LLMs raise the transfer coefficient, converting tacit fragments into new explicit forms of automated processing.

Nevertheless, the scope has limits. Although LLMs model aspects of tacit knowledge, they do not experience the world. They have no body, emotions, or direct physical perception. Their knowledge is a symbolic projection, not a lived understanding. And it is here that Polanyi’s paradox remains powerful: although we can codify many aspects of human experience, not all understanding can be encapsulated in patterns of data.

Thus, the development of LLMs and NLP can be seen as a partial force of erosion against tacit knowledge: powerful for certain cognitive tasks, but limited in everything that depends on living, feeling, and deciding in immediate physical and social worlds.

In this emerging landscape, the real challenge will not only be to build better models, but to understand which parts of human knowledge must remain outside the algorithms.

Automation, Complementarity, and Job Polarization

At the heart of this analysis lies the question of how the transformation of tacit into explicit knowledge impacts employment.

When a task becomes explicit, it opens the door to automation. The greater the explicit value of a task, the higher its automation potential, measured in our model as A. This implies that certain jobs once considered safe could become replicable by machines if tacit knowledge transfer becomes efficient.

However, as long as there remains non-transferable tacit knowledge, there will be room for human complementarity, represented in our model as C. Complementarity occurs when human presence adds value to processes that cannot be fully described or predicted.

The tension between automation and complementarity defines the net displacement of employment, which in our model depends on the weighted difference between A and C.

What emerges is a phenomenon already observed empirically in advanced economies: job polarization.

High-skill jobs requiring creativity, flexibility, and judgment tend to grow. Some low-skill jobs, especially those requiring physical interaction or adaptability, also resist or evolve.

In contrast, middle-skill jobs, based on routine and structured tasks, tend to disappear under the pressure of automation.

Our net displacement graphs illustrate this dynamic: as the transfer of tacit knowledge (via J) becomes more effective, the risk of displacement increases, although unevenly.

Dynamic Visualizations

To illustrate these ideas, we developed a series of visualizations:

Show original animation

The second visualization illustrates how, as the erosion of tacit knowledge progresses, the possibility of human complementarity decreases. This occurs because tasks that required judgment, adaptability, or implicit interpretation become increasingly codifiable and predictable.

Show original animation

A third visualization integrates both effects—automation and complementarity—to calculate the net displacement of employment across different technological advancement scenarios. This three-dimensional surface shows how different labor sectors are affected unevenly.

Show original animation

Finally, we simulated how the knowledge structure of tasks evolves dynamically under different values of physical interaction and real-world learning.

With a moderate J value (0.5), automation advances but there is still considerable space for human complementarity.

Show original animation

With a high J value (0.9), however, automation almost completely dominates, leaving minimal room for human intervention.

Show original animation

These visualizations make it possible not only to understand the theory but also to tangibly see how tacit knowledge and its erosion redefine the boundaries between humans and machines.

Final Reflections

Technology does not advance in a vacuum. It evolves by interacting with human knowledge, adapting it, and sometimes transforming it.

Polanyi’s paradox remains valid: even in an age of machine learning, there is a core of knowledge that resists codification.

Experience, tacit knowledge, and common sense continue to be our greatest differential.

The future of work will increasingly be a balance between what machines can automate and what humans can still do better, not because we say so, but because we know it, in ways we still cannot program.