Tacit knowledge and Generative AI

AIFuture of Work
Written By Andrew Tope

A white paper published by Intel recently makes a remarkable claim. The researchers at Intel, working with the U.S. Census Bureau, say they have built a system that captures "tacit knowledge" from subject matter experts and scales it through artificial intelligence.

The system, called a "Context Atlas," works through feedback loops. Experts review AI outputs, correct errors, and those corrections are stored and injected into future prompts. Over time, the system accumulates a repository of expert patterns. The reported results are striking: processing time dropped from weeks to minutes, hallucination rates fell by 40%, and variation between different experts' outputs dropped by 99.9%. "Institutional knowledge," the authors write, "was transformed into a structured resource usable by other teams at scale."

The language is confident. Tacit knowledge, the paper explains, is "insight gained through experience that experts rely on to interpret complex and ambiguous data." The Context Atlas captures this insight and deploys it without costly model retraining. The implication is clear: what once lived only in experts' heads can now be extracted, stored, and scaled.

There is just one problem. Tacit knowledge, by definition, cannot be captured. That is what makes it tacit.

What We Know But Cannot Tell

Michael Polanyi (the Hungarian-British polymath whose work on knowledge spanned physical chemistry, economics, and philosophy) articulated the problem in a phrase that has echoed through six decades of research: "We know more than we can tell."

Writing in The Tacit Dimension in 1966, Polanyi observed that human knowledge comes in two fundamentally different forms. Explicit knowledge can be codified, written down, transmitted in manuals and procedures. You can explain the rules of chess, the steps of a recipe, the formula for compound interest. Tacit knowledge cannot be articulated this way. You know how to ride a bicycle, but try writing instructions that would enable someone to learn without getting on one. You recognise your mother's face in a crowd, but try specifying the algorithm. You know when a sentence sounds wrong, but good luck articulating the grammatical rule you're applying.

The distinction matters because capture requires codification. To store knowledge in a system, you must first express it in a form the system can hold: rules, patterns, corrections, examples. Tasks rich in tacit knowledge resist this process. The knowledge exists in the practitioner's body, their intuitions, their feel for the work. It cannot be extracted and transferred to a machine because it cannot be extracted and transferred at all, even to another human, except through prolonged apprenticeship and practice.

This is not a limitation of current technology. It is a feature of what tacit knowledge is.

The Putter-Togetherer's Hands

Ian Leslie, writing recently in his newsletter The Ruffian, explored a case that illuminates what tacit knowledge actually looks like in practice. He describes a visit, recounted in James Fox's book Craftland, to Ernest Wright, one of the last scissor manufacturers in Sheffield, England.

A quality pair of scissors involves at least seventy discrete manufacturing processes. But the most critical step belongs to the "putter-togetherers": craftspeople who connect and align each pair of blades. Why is this role so important? Because each pair of blades is unique. As one craftsman explains: "A correctly assembled pair of scissors needs its bows aligned, its blades curved, its points crossing, the cutting tension firm and even all the way along."

The crucial line: "No machines can handle that. You can't even learn it in a book."

This is tacit knowledge in action. The putter-togetherer has developed, over years of practice, a feel for how a particular pair of blades needs to be adjusted. The problem is never the same twice. Each represents what Leslie calls "a singular problem demanding a bespoke solution." The knowledge required to solve it cannot be written down because it was never propositional in the first place. It emerged from practice and exists only in practice.

Now consider what the Context Atlas would capture from this craftsman. It could record his corrections: "When the blades gap at the tip, adjust the tension here." It could accumulate hundreds of such corrections until a pattern library emerged. But could it capture the feel that recognises when this particular pair of blades requires something the system has never seen? Could it capture the judgment that knows when the standard adjustments won't work?

It could not. Because that judgment is not a correction to an output. It is a way of seeing that cannot be reduced to propositions.

The Category Error

The white paper conflates several distinct types of knowledge, and the conflation matters.

Domain knowledge consists of facts about a field: regulatory requirements, classification schemas, institutional procedures. The Census Bureau has rules for categorising business survey responses. Those rules can be documented, stored, and injected into prompts. This is explicit knowledge that happens to be undocumented.

Processing rules specify how to handle specific cases: if a response mentions a subsidiary, check for parent company information. These too are explicit and codifiable. Experts may carry them in their heads rather than in manuals, but they can be articulated when prompted.

Tacit judgment is something different. It is the ability to recognise that something is wrong without being able to specify what. It is knowing when the standard rules don't apply. It is the feel that flags a case as unusual before any explicit reasoning has occurred. This, by definition, cannot be articulated. If it could be articulated, it would be explicit knowledge.

The Context Atlas captures the first two effectively. When an expert corrects an AI output, they make their judgment explicit. The correction itself becomes codifiable. The system surfaces knowledge that was merely undocumented and makes it available at scale. This is genuinely valuable.

But the paper claims to capture tacit knowledge. This is a category error. What the system captures are explicit corrections that experts generate when reviewing outputs. The tacit judgment that enabled those corrections, the feel that knew something was wrong, remains in the expert's head. It has to. That is what tacit means.

The Consistency Trap

Consider the paper's most striking metric: "SME-to-SME variation in GenAI outputs dropped by 99.9%."

On its face, this sounds like an achievement. Consistent results regardless of which expert provides oversight. Standardisation, reproducibility, scale.

But think about what this means. In domains with genuine tacit knowledge, experts often disagree. Two experienced diagnosticians read the same case differently. Two seasoned account managers handle the same client with different approaches. This variation is not noise to be eliminated. It often reflects the irreducible ambiguity of the domain, the legitimate space for different expert judgments, the multiple valid ways of exercising tacit skill.

When the Context Atlas reduces variation to near zero, it has standardised outputs. But standardised according to whom? The accumulated corrections of whichever experts happened to review the training cases. The system has achieved consistency by eliminating the perspectives of experts who would have judged differently.

In rule-governed domains with single correct answers, this is appropriate. Census data classification likely has right answers that experts would agree on. But the paper claims the system applies across "government, finance, law, and healthcare." In domains with genuine ambiguity, the 99.9% consistency may be a bug, not a feature. It may mean the system has learned one expert's perspective while losing the judgment that would have flagged when that perspective fails.

What History Teaches

The pattern is not new. In the 1970s and 1980s, the expert systems movement made strikingly similar promises. MYCIN, developed at Stanford, encoded the diagnostic rules of infectious disease specialists into a knowledge base of if-then statements. It achieved 65% accuracy, matching many specialists. Researchers proclaimed a revolution: expert knowledge could be captured and scaled.

By the early 1990s, most expert systems projects had failed. Not because they didn't work in controlled settings. Because they worked only within the boundaries of what had been explicitly encoded. Novel cases, edge cases, cases that combined familiar elements in unfamiliar ways: the systems failed precisely where expert judgment mattered most.

Hubert Dreyfus (the philosopher at Berkeley whose critique of artificial intelligence proved prescient) had predicted this. "The problem," Dreyfus argued, "is that experts don't follow rules. They have intuitions based on experience that they cannot articulate." The expert systems captured the explicit layer of expertise while missing the tacit layer underneath. The rules worked until they encountered a situation the rules didn't cover. And the system had no way of knowing it had encountered such a situation.

The Context Atlas is architecturally more sophisticated. It doesn't require knowledge engineers to laboriously extract rules. It captures corrections in the natural flow of work. It combines accumulated patterns with the general capabilities of large language models. These are genuine improvements.

But the fundamental limit remains. The system captures what experts make explicit. It cannot capture what experts cannot articulate. And the most valuable expert judgment, the judgment that knows when patterns fail, often falls into that second category.

The Novice's Danger

Harry Collins (the sociologist of science at Cardiff University whose taxonomy of tacit knowledge has become standard) distinguishes between knowledge that is tacit contingently and knowledge that is tacit necessarily. Contingently tacit knowledge is merely undocumented: no one has written it down, but in principle someone could. Necessarily tacit knowledge cannot be articulated because it depends on embodied experience and contextual judgment that resists propositional form.

The Context Atlas excels at surfacing contingently tacit knowledge. When experts correct outputs, they document what was previously undocumented. The system makes explicit what was merely implicit. This creates genuine value, particularly in organisations where institutional knowledge lives only in experienced staff.

But necessarily tacit knowledge resists this process by its nature. And here is the danger: a novice using the system has no way of knowing the difference.

The novice submits a query. The system returns a response enriched with accumulated expert corrections. The response handles known patterns correctly. But what happens when the novice encounters a case that doesn't match the patterns? They have no way of knowing that this case is different. The system gives them confidence without the judgment that would calibrate it.

The expert systems movement discovered this problem the hard way. MYCIN handled common infections well, but doctors using it couldn't tell when they had encountered an uncommon case where the rules didn't apply. The system created what researchers called "proficient novices": people who could apply patterns competently but lacked the judgment to know when patterns failed.

The Context Atlas risks the same outcome. It can make novices perform better on standard cases while leaving them exposed on non-standard ones, and giving them no signal that they're exposed.

What the System Actually Does

To be clear: the Context Atlas likely delivers real value. Processing time reductions of 240x and 40% fewer hallucinations represent genuine improvements. Surfacing undocumented institutional knowledge and making it available at scale solves a real problem. Organisations lose enormous amounts of practical knowledge when experienced staff leave; systems that capture even the explicit portion of that knowledge are worth building.

The issue is not that the system is worthless. The issue is that "capturing tacit knowledge" overstates what it achieves. The system captures explicit corrections. It surfaces contingently tacit knowledge. It standardises handling of known patterns. These are valuable but different from capturing the judgment that knows when patterns fail.

The distinction matters because it affects how the system should be deployed. In domains characterised by explicit rules and verifiable answers, the Context Atlas likely performs as advertised. In domains characterised by ambiguity, judgment, and the need to recognise genuinely novel cases, it should be treated with more caution, and its outputs should be overseen by people with the tacit judgment it cannot contain.

Practical Implications

What does this mean for practitioners evaluating or deploying knowledge capture systems?

  • Audit your domain before deploying. Ask: how much of our experts' value comes from applying documentable rules versus exercising judgment that resists articulation? The more rule-governed the domain (compliance checking, data classification, routine document review), the better these systems will work. The more judgment-dependent (complex diagnosis, negotiation, strategic decisions under ambiguity), the more caution is warranted.

  • Distinguish efficiency gains from judgment capture. The Context Atlas likely delivers real efficiency: faster processing, fewer obvious errors, more consistent handling of standard cases. These gains are valuable. But they are not the same as capturing expert judgment. Celebrate the first without claiming the second.

  • Monitor for edge case blindness. The critical question is not how the system handles typical cases but how it signals atypical ones. A system that handles 95% of cases correctly while giving no warning on the remaining 5% may be more dangerous than one that handles 80% correctly but flags uncertainty. Ask your vendor: what does the system do when it encounters something outside the accumulated corrections?

  • Preserve apprenticeship pathways. If the system creates the illusion that expert knowledge has been captured, organisations may underinvest in developing new experts. But tacit knowledge transfers only through practice and proximity. The Context Atlas might preserve some institutional knowledge, but it cannot replace the development of human judgment. Protect your mentorship pipelines.

  • Read the claims carefully. "Capturing tacit knowledge" is a strong claim. "Capturing expert corrections and surfacing undocumented institutional rules" is a different, more defensible claim. The gap between them is where deployment risk lives. Press vendors on which claim they can actually support.

The Irreducible Human

The putter-togetherer at Ernest Wright has spent years developing a feel for how particular blades need to be aligned. That feel exists in his hands, his eyes, his accumulated experience of thousands of unique problems demanding bespoke solutions. It cannot be written down because it was never explicit in the first place.

Every generation of technologists rediscovers this limit. Expert systems tried to capture medical diagnosis in the 1970s. Knowledge management systems tried to capture organisational wisdom in the 1990s. The Context Atlas tries to capture tacit knowledge through feedback loops today. Each system captures something valuable: explicit rules, documented procedures, accumulated corrections. Each system stops at the same boundary: the judgment that cannot be articulated because it was never propositional.

Polanyi understood this sixty years ago. The craftsmen at Ernest Wright understand it in their hands. What we know most deeply, we often cannot tell. That knowledge lives in practice, transfers through apprenticeship, and resists every attempt at extraction.

The Context Atlas is a useful tool. But tacit knowledge remains uncapturable. That is not a problem to be solved. It is a feature of what human expertise actually is.

Andrew Tope
Previous

AI Transformation and the McNamara fallacy
Next

The pattern matching trap