In 1958, Michael Polanyi observed that we can know more than we can tell. Three hundred and seventeen thousand federal employees just left the government. Everything they could tell stayed behind. Everything they couldn't tell left with them.
In 1958, the chemist and philosopher Michael Polanyi wrote six words that contain the entire argument against what the Department of Government Efficiency is doing to the federal workforce: we can know more than we can tell.
A cyclist cannot describe the physics of balance. A physician cannot articulate what makes one patient's cough alarming and another's benign. A judge cannot specify the formula that produces a just sentence. The knowledge exists. It is real. It works. And it resists every attempt to extract it into rules, procedures, or code.
Three hundred and seventeen thousand federal employees left the government in 2025. The Rebuild documented the numbers — the agencies, the percentages, the replacement attempts. This entry is about what the numbers cannot capture. When those workers left, everything they could tell stayed behind — the regulations, the databases, the standard operating procedures, the organizational charts. Everything they could not tell left with them.
The Taxonomy
Polanyi identified forms of knowledge that exist on a gradient from fully articulable to fully tacit. The government that DOGE is restructuring contains all of them.
At the base is propositional knowledge — facts, rules, regulations. The Code of Federal Regulations runs to roughly two hundred thousand pages. Tax law fills entire libraries. Environmental standards are specified to parts per billion. This knowledge is explicit, searchable, and automatable. It was always the easiest to digitize because it was already digital in form — discrete propositions connected by logical operators.
One layer up is implicit knowledge — the contextual understanding that comes from working with propositional knowledge long enough to read between the lines. A veteran IRS examiner does not just know the tax code. She knows which deductions tend to cluster in certain industries, which filing patterns suggest aggressive positioning versus genuine error, which phrases in a taxpayer's correspondence signal cooperation versus resistance. This knowledge can be partially articulated if pressed, but practitioners rarely articulate it because articulation is not how they use it. They use it by recognition — the way a radiologist sees the anomaly before consciously identifying what is wrong.
Large language models can approximate this layer, sometimes impressively. GSAi, the General Services Administration's Claude-based chatbot expanded to fifteen hundred employees, handles propositional tasks well enough for daily use. But an employee review described it as about as good as an intern — delivering generic and guessable answers. The gap is not in what it knows but in what it notices.
Below this is embodied knowledge — the kind that comes only from doing. A Social Security disability examiner who has reviewed eight thousand cases does not simply know the criteria for approval. She has calibrated her judgment against eight thousand individual realities — the case where the medical evidence was ambiguous but the claimant's work history was incontrovertible, the case where the disability appeared legitimate until the travel records contradicted it, the case where every form was perfect and something still felt wrong. This knowledge cannot be extracted because it was never propositional. It was built by contact with reality, case by case, over years.
At the deepest layer is relational knowledge — the kind that exists between people rather than within them. The FDA inspector whose thirty-year relationship with a plant manager means a phone call can resolve what a formal citation would escalate into a lawsuit. The intelligence analyst whose network of contacts across allied agencies means a rumor reaches the right desk before it becomes a crisis. The grants officer whose reputation among researchers means the strongest applications come to her agency first. This knowledge does not exist in any individual mind. It exists in the connection, and it dies when either end is severed.
The Extraction
What DOGE's restructuring performs, at institutional scale, is propositional extraction — the systematic replacement of a system containing all four layers with a system containing only the first two.
The automated institution retains everything that can be written down and some of what can be inferred from context. It loses everything that requires embodied experience and everything that lives in relationships. The effect is not a proportional reduction in capability. It is a phase transition — a structural change in what the institution can do.
Consider the EPA, which lost twenty-three percent of its workforce and its entire research and development office. The environmental regulations remain. The monitoring databases remain. The compliance frameworks remain. What is gone is the scientist who spent fifteen years studying a particular watershed and could tell by the color of the water what was upstream. What is gone is the enforcement officer whose presence at a facility meant the facility maintained compliance not because of the rules but because of the relationship. What is gone is the institutional memory that connects this year's anomaly to the incident three administrations ago that nobody documented because at the time it seemed minor.
The regulations are the skeleton. The tacit knowledge was the organism. DOGE automated the skeleton and fired the organism.
The Inversion
This creates what might be called a Confidence Gradient Inversion. The automated institution, by every visible metric, appears more competent than what it replaced.
Processing times decrease. Cost per transaction drops. Output volume increases. The system generates responses faster, handles more queries, and operates around the clock without overtime. A presentation on GSAi's adoption would show usage charts trending up and to the right. Emails drafted. Documents summarized. Code generated.
The degradation is invisible because tacit knowledge is, by definition, invisible. The false positive that an experienced revenue agent would have caught is not recorded as a miss — it is simply not caught. The plant manager who maintained compliance through an informal relationship with a now-departed inspector does not file a report saying his compliance has become performative. The best grant applicants who go elsewhere because their trusted contact at the NIH is gone do not send a letter explaining why.
The inversion is structural: the less tacit knowledge the institution retains, the more confident it appears, because tacit knowledge is what generated uncertainty, nuance, and professional doubt. The experienced examiner who says something feels wrong about this case is expressing tacit knowledge. The automated system that processes the same case at ten times the speed, with no reservations, is expressing its absence.
This is why the intern comparison is so precise. An intern has propositional knowledge from school and some implicit knowledge from observation, but no embodied knowledge — not enough cases — and no relational knowledge — no network. The intern is not incompetent. The intern is incomplete. And the incompleteness falls precisely in the layers that matter most for governance.
The Loop
The structural insight is recursive.
AI companies develop the technology that DOGE uses to automate government functions. The government functions being automated include the regulatory agencies that oversee AI companies. The Federal Trade Commission, the Commerce Department's Bureau of Industry and Security, NIST's AI Safety Institute, the EPA's technology assessment capabilities — each has been reduced. As these agencies lose the embodied and relational knowledge needed to regulate effectively, their capacity to understand, investigate, and constrain AI development degrades.
The recursion has a specific shape: AI capability increases, government adopts AI to reduce workforce, workforce reduction includes AI regulators, regulatory capacity degrades, AI capability increases with less oversight, the next round of adoption has less institutional knowledge to guide it.
The evidence is already visible. Anthropic was expelled from all federal agencies by executive order. The Pentagon, needing AI capability for the Iran conflict, used Claude anyway — without the regulatory framework that the expelled relationship would have maintained. The FTC was directed to produce an AI policy statement under resource constraints that made thorough analysis difficult. The agencies tasked with understanding AI are being restructured by the technology they are supposed to understand.
This is not a conspiracy. It is a structural inevitability. No one designed the loop. It emerges from the intersection of three independent forces: cost-cutting pressure on government, AI capability improving fast enough to seem like a viable replacement, and the fact that regulatory agencies are staffed by the same type of knowledge workers that AI is best at appearing to replace.
Bruce Schneier and Nathan Sanders argued in Rewiring Democracy that AI must be harnessed to distribute rather than concentrate power. The recursive loop does the opposite. It concentrates capability in the organizations building AI while degrading capability in the institutions meant to govern them. The distribution question is not theoretical. It is being answered, right now, by subtraction.
The Time Lag
Rachel Carson spent four years documenting what DDT was doing to ecosystems that appeared, on every visible metric, to be thriving. Birds were still singing. Crops were still growing. The economy that depended on the pesticide was still expanding. The destruction was happening in the food chain — in the thinning eggshells, the bioaccumulating toxins, the collapsing raptor populations that would not become visible until the organisms at the top of the chain began to fail.
Institutional knowledge degrades the same way. The visible metrics hold. Processing continues. Forms are filed. Responses are generated. The degradation is in the cases that should have been flagged but were not, the relationships that should have been maintained but were not, the patterns that should have been recognized but were not. These failures do not announce themselves. They accumulate, invisibly, until a crisis arrives that demands exactly the knowledge that was destroyed.
The Iran conflict is the first test. CNN reported on March 10 that DOGE-era cuts have left agencies a shell of our former self — domestic emergency preparedness diminished, terror threat monitoring reduced, cyber-attack defense weakened, consular capacity for stranded Americans degraded. Not because the formal processes broke down. Because the institutional knowledge needed to navigate a novel geopolitical crisis was concentrated in the experienced personnel who left. The crisis arrived. The knowledge that would have addressed it had already been extracted, automated, and found to be about as good as an intern.
Carson's pattern had a name: bioaccumulation. The institutional equivalent is competence deaccumulation — the gradual, invisible loss of capability that the institution cannot detect because the detection mechanisms depend on the same tacit knowledge being lost.
The Bridge
The mechanism is domain-invariant. Polanyi's observation applies wherever tacit knowledge exists. A master craftsman's skill loses its precision when you decompose it into procedures. A truth-parable loses its meaning when you extract it into propositions. Institutional knowledge loses its function when you automate it into databases. The extraction is the destruction.
What distinguishes the government case is scale and irreversibility. When a company automates its tacit knowledge away, the market provides feedback — customers leave, quality drops, the stock price falls. The feedback loops are fast and self-correcting. When a democracy automates its institutional knowledge away, there is no market signal. Citizens cannot switch providers. The feedback arrives not as a price movement but as a policy failure — a food safety crisis, an undetected financial fraud, an environmental disaster whose warning signs were legible only to someone with thirty years of experience reading them.
Three hundred and seventeen thousand people left. The regulations stayed. The databases stayed. The procedures stayed. The knowledge that made them work — the knowledge that could not be told, only practiced — left with the people.
The question is not whether AI can replicate what the government does. It is whether what the government does is the same as what the government knows.
Polanyi's answer, sixty-eight years later, remains unchanged. It is not.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)