Can AI Learn What Humans Can’t Explain? The Tacit Knowledge Problem

What Is Tacit Knowledge?

Decades before deep learning became a household term, the philosopher Michael Polanyi captured an essential feature of human expertise: “We can know more than we can tell.” Riding a bicycle, recognizing a face, or crafting a persuasive argument all involve knowledge that is deeply embedded in practice, intuition, and context. This is tacit knowledge—the kind of skill people deploy masterfully but often struggle to articulate, formalize, or transfer as a set of explicit rules.

Unlike explicit knowledge, which can be written down in manuals or encoded in databases, tacit knowledge lives in the body, in the subtle cues of a social interaction, or in the unspoken heuristics that guide a seasoned professional. For cognitive scientists and AI researchers alike, a pressing question now resurfaces: is this type of knowledge merely difficult to capture with today’s algorithms, or is it fundamentally inaccessible to any machine?

AI’s Struggle with Generalization

Modern artificial intelligence systems are, in many respects, narrowly superhuman. They defeat world champions in Go, generate fluent text, and fold proteins with astonishing accuracy. Yet they remain brittle outside their training distributions. A chess engine cannot hold a casual conversation; a language model may fail at simple arithmetic when the problem is phrased unconventionally. The latest analyses in venues such as the Communications of the ACM frame this brittleness as a symptom of missing tacit knowledge. The argument is that today’s narrow AI systems, no matter how large, lack the rich, context-sensitive understanding that humans acquire through embodied experience and social learning.

A narrow AI can excel at a task whose rules are fully specifiable in the training data—language patterns, game moves, protein structures—but it often fails to adapt when the task shifts even slightly. This deficit is more than a performance gap; it points to a cognitive style that does not generalize in the fluid, cross-domain way that human intelligence does.

Is Tacit Knowledge Encodable—or Inaccessible?

The debate is as much philosophical as technical. On one side, optimists argue that tacit knowledge is simply a data problem. Given enough examples—video of humans navigating social situations, sensor logs from physical tasks—a sufficiently powerful model might infer the unwritten rules. Researchers in robotics and reinforcement learning are already attempting to capture practical know-how through massive interaction datasets. In this view, tacit knowledge is “hard to encode” but not impossible, and future architectures may bridge the gap.

On the other side, skeptics point to foundational limits. Symbolic reasoning systems of the past failed to capture the fluidity of human insight precisely because they demanded full explicitness. If some knowledge is inherently non-propositional—experienced but not describable—then no amount of pattern matching on past data can internalize it. As the Stanford Encyclopedia of Philosophy entries on tacit knowing note, the very act of bringing certain skills under verbal description can distort them, a phenomenon that would equally constrain any explicit representation inside a machine.

The tacit dimension is not a bug in human cognition but a feature of embodied, situated intelligence—a feature that current AI architectures may never duplicate by mere scaling.

Implications for AI Safety and Progress

If tacit knowledge is indeed irreplaceable, then the road to robust general intelligence demands more than bigger models and more data. It may require fundamentally different machine architectures that incorporate embodiment, interaction, and even aspects of consciousness that cognitive science is still mapping. For the AI safety field, this is a double-edged sword: it might slow the arrival of superintelligent systems that could pose existential risks, but if such systems do emerge without real-world common sense, their decisions could be dangerously alien.

The Communications of the ACM piece does not hand down a verdict. Instead, it asks the community to take the possibility of irreducible tacit knowledge seriously as a scientific question. By doing so, it connects contemporary AI research to enduring puzzles about the nature of mind and expertise—reminding us that the deepest challenges to machine intelligence may not be computational, but conceptual.