Tacit Knowledge Limits AI Coding Assistants

The essay published on cekrem.github.io (May 19, 2026) revisits Michael Polanyi's 1966 central claim, "We can know more than we can tell." The author reports Polanyi's argument that much expert knowledge is tacit, it lives in practice and embodied pattern recognition and cannot simply be written down or transcribed into a dataset. The essay illustrates the point with software examples, noting senior engineers who instantly recognise a problematic pull request and only later can articulate why, and with analogies to surgeons and experimental scientists. The piece frames this as a structural barrier for contemporary AI coding assistants, because the knowledge to learn from is not present in textual corpora.

From an industry-technical perspective, tacit knowledge maps to skills that are learned via repeated interaction, feedback loops, and embodied judgement rather than explicit rules or documentation. Model training relies on surface signals in code and text. Industry-pattern observations: models excel where the signal is codified and repeatable but struggle where competence depends on long-run, context-rich feedback and situated practice.

For practitioners, the essay highlights a recurrent pattern: augmentative AI is strongest at automating routine, well-specified tasks and surface-level refactoring, while high-stakes code design, architectural judgement, and mentorship-style knowledge transfer remain human-centred activities. This framing affects expectations for code-review automation, pair-programming assistants, and the metrics teams use to evaluate tooling.

Indicators worth tracking include whether tooling prioritises signals that proxy practice and whether evaluation benchmarks move beyond static code corpora to include temporally extended, apprenticeship-like feedback. Observers should also watch for research that attempts to formalise or approximate tacit signals without asserting access to the original embodied knowledge.