Tacit Knowledge in the Era of AI

One thing I always remind myself, whether I'm in a heated discussion with my parents or debating something with friends, is to look from both sides. Put myself in the other person's shoes. It's a simple habit, but it changes everything about how you understand a situation.

So I did it with AI.

Everyone's talking about how fast AI is moving. How it'll surpass us someday. And there are already things AI can do that humans simply cannot: process millions of data points, never forget, work without sleep. When we discuss all of this, we do it from our perspective, as humans. That's natural. We're the ones who feel threatened, inspired, or curious. We are humans after all, and we are not AI. So we talk, we discuss, we form opinions from where we stand.

But what if we flipped it?

What if I put myself in AI's shoes? If I were an AI, trying to surpass human intelligence, how would I think about it? Where would I see the gap?

First, I'd take stock of what I have. And the inventory is impressive. I've consumed virtually every book ever digitized. Every Wikipedia article. Every Stack Overflow answer. Billions of lines of code. Academic papers, legal filings, medical records, news archives, YouTube transcripts. Images. Videos. If it was ever uploaded to the internet, I've probably seen it.

So what's the gap?

Here's where it gets uncomfortable. Not for humans, but for the AI. Humans don't write everything down.

Michael Polanyi, a philosopher who spent decades thinking about the structure of knowledge, put it simply in his 1966 book The Tacit Dimension: “We know more than we can tell.” He called this tacit knowledge, the kind that lives in your body, your intuitions, your reflexes. The senior engineer who “smells” a bad architecture before they can explain why. The product manager who knows which customer feedback to ignore. The chef who adjusts a recipe by feel, not measurement.

In 2014, MIT economist David Autor gave this idea a name in the economics literature: Polanyi's Paradox. He argued that “the tasks that have proved most vexing to automate are those demanding flexibility, judgment, and common sense, skills that we understand only tacitly.” Decades later, this still holds.

The numbers back it up. A study by the Delphi Group found that 42 percent of corporate knowledge resides solely in employees' heads. Not in documents. Not in databases. In people. Nonaka and Takeuchi, who built the foundational model of organizational knowledge creation, used the iceberg metaphor: explicit knowledge is just the tip above the waterline. Knowledge management researchers across multiple studies consistently estimate that 80 to 90 percent of what an organization actually knows is tacit, embedded in employees' experiences, instincts, and skills.

As an AI, this is my ceiling. I can master the explicit layer. I can even push that boundary, connecting dots between documents no single human could read in a lifetime. But the vast majority of human knowledge? It was never on my training data. It was never anywhere I could reach. Even a 2025 paper in the Review of Austrian Economics examining whether large language models possess tacit knowledge concluded that while LLMs can exhibit some forms of hard-to-codify knowledge, they cannot replicate embodied knowledge gained through sensory experience, and remain subject to what Hayek called the constraints of dispersed knowledge. If I cannot access the tacit layer, I cannot surpass even my own ceiling, let alone match the full depth of a human.

The Nervous Generation

This is not just a theoretical argument for me. I have been running a non-profit called Project X Vietnam for three years now, and every year I face a new cohort of fellows. Young, talented people who are really active in the early stages of their careers, across product, engineering, growth, design, and many other aspects of business. They are sharp. They are hungry.

And they all have the same question: what is the point of building a career when AI is changing everything this fast?

Many of them are optimistic about their own skills. They know they are capable. But they are pessimistic about the pace. Entry-level roles are shrinking. The bar for seniority keeps climbing. The middle is getting squeezed. They look at the landscape and think, if AI can do the junior work, and companies only want seniors, then how do I ever cross that gap?

The anxiety is real. I see it every year, in every batch. And it is getting worse, not better.

Why Senior Roles Are Rising and Entry Roles Are Shrinking

Here is the thing that most people feel but few articulate clearly: the shift in hiring is not random. It maps directly onto the tacit knowledge argument.

Think about what a junior employee primarily does. They execute well-defined tasks. They write code from specs. They build screens from designs. They research topics and summarize findings. They draft documents. They follow processes. Most of this work operates on the explicit layer: take clear input, produce clear output.

AI is now very good at exactly that. It can write code from a spec. It can generate designs. It can summarize research. It can draft any document you need. When a company looks at the cost of a junior hire versus the cost of an AI subscription, the math is brutal. Not because juniors are bad, but because the explicit layer of their work is the exact layer AI has mastered.

Now think about what a senior employee does. Yes, they can do all the junior tasks faster. But that is not why companies pay them two or three times more. Companies pay for the tacit layer.

A senior product manager does not just write a PRD. They know which PRD to write. They walk into a room full of stakeholders with competing priorities and walk out with alignment, not because they have a framework for it, but because they have been in that room fifty times before and they can read the dynamics. They know that the VP of Sales says they want feature X, but what they actually need is a better onboarding flow. They know this not because anyone told them, but because they have seen the same pattern three companies ago and watched it play out.

A senior engineer does not just ship code. They know which code not to write. They look at a feature request and think, “if we build this the obvious way, it will create a migration nightmare in eight months when we need to support multi-tenancy. ” Nobody taught them this in a course. They learned it by living through the migration nightmare at a previous company. That scar tissue is worth more than any amount of clean code an AI can generate.

A senior designer does not just make things look good. They know why a certain layout will confuse users before any user test confirms it. They have internalized thousands of hours of watching real people struggle with interfaces. They can look at a mockup and feel, physically feel, where the friction is. That is not a skill you can prompt for.

A senior growth lead does not just run experiments. They know which experiments are worth running. They have a sense for which channels are about to get saturated. They have seen three different companies burn through the same paid acquisition playbook and know exactly when it will stop working. They carry a mental model of how markets shift that no dashboard can replicate.

This is why the demand curve has shifted. Companies are not paying for hands that execute. They are paying for brains that have absorbed years of context, failure, and judgment. They are paying for tacit knowledge. And only companies with enough budget to invest in long-term human development are still hiring juniors, because they understand that today's junior, given the right environment, becomes tomorrow's senior. The rest of the market has decided that AI can handle the explicit layer cheaply enough that they would rather hire fewer people who already carry the tacit layer.

It is a rational decision at the company level. But it is a terrifying one at the individual level, especially if you are 22 and trying to start.

What I Tell Them

But there is one thing I always tell my fellows. AI will not easily replace you, especially once you start holding tacit knowledge. And tacit knowledge is something you begin accumulating the moment you start doing real work. Not reading about it. Not watching tutorials about it. Actually doing it. Building things. Making decisions. Failing. Iterating. Feeling the texture of a problem that no documentation can capture.

The entry-level squeeze is real, but it does not mean the path is closed. It means the path runs through doing, not through credentials. Through apprenticeship, not through courses. Through proximity to people who carry tacit knowledge and are willing to let you absorb it. That is exactly what we try to create at Project X Vietnam: not a classroom, but a context where tacit transfer can happen.

Domain Knowledge Runs Deep

Let me make this concrete. I work in proptech right now. I build property management software. After spending real time in that domain, you start to understand things that are nowhere in any manual. You understand why landlords in Belgium behave differently from landlords in the Netherlands. You understand the subtle friction points in how property managers communicate with tenants. You know which features sound great on paper but will never get adopted, because you've watched the adoption curve fail three times before.

When a new product challenge comes up in property management, I don't start from a Google search. My brain activates patterns from past failures, past conversations, past decisions that I made two years ago and barely remember consciously. Something just surfaces. A feeling of “this is similar to that thing we tried in 2024 that didn't work because of X.” And X is not written down anywhere. It lives in me.

An AI cannot replicate this. It can read every proptech blog, every case study, every product teardown ever published. But the transferable, domain-specific experience of actually having built and shipped and failed in that space? That is not text. That is not data. That is tacit knowledge, and it is mine.

Same Book, Different Paths

Here is a thought experiment that I keep coming back to.

Take a human and an AI. Give them the same book on product management. They both read it cover to cover. At that point, they share the exact same explicit knowledge. They know the same frameworks, the same terminology, the same case studies.

Now give them the same assignment: build something with that knowledge.

The paths diverge immediately.

The AI will reference best practices. It will combine patterns from what it has seen work in other contexts. It will produce something that looks polished, structured, well-reasoned. But it does this without any understanding of why. It is, at its core, text generation. Sophisticated text generation, but text generation nonetheless. Everything it does is a cooking mixture: combination, reuse, recycling of explicit knowledge that exists out there.

The human does something different entirely.

The human asks questions. Thinks of some ideas. Takes one idea and goes to talk to people about it. Iterates. Starts to form a conviction. Builds something. It fails. They understand why. They iterate again. And at every single point in that process, something happens that cannot be put into words. It feels like something. Not in a mystical sense, but in the sense that the experience of navigating ambiguity and making decisions under uncertainty generates a kind of knowledge that is fundamentally different from anything that can be written down.

And here is the question that really stops me: throughout that entire process, how much of it does anyone actually document? How many of us, how many of even myself, write down the full chain of thinking that led to a decision? The dead ends we explored. The conversations that shifted our perspective by 10 degrees. The moment when something just clicked and we changed direction for reasons we could barely articulate even to ourselves.

We don't document it. We can't document it. And even if we could, the documentation would be a pale shadow of the actual experience. The AI would read our notes and get the what, but never the how it felt to arrive there.

The Intuition Problem

This brings me to something I think about a lot: intuition.

We like to pretend that our decisions are purely logical. Data driven. Backed up by evidence. And the good ones often are, to some degree. But even the most data-driven decision has an intuitive core. Choosing to open a franchise restaurant based on market research, demographic data, location analysis, financial models? That decision is still, at its foundation, a leap. You trust that the data will support your success. You trust your reading of the data. You trust your judgment about which data matters and which to ignore.

If decisions were purely logical and every well-backed decision guaranteed success, then everyone in the world with access to the same data would succeed equally. But they don't. The gap between data and decision is intuition, and intuition is built from the accumulation of every experience you have ever had, most of which you have forgotten consciously but which your brain has encoded as pattern, as instinct, as a subtle pull toward one option over another.

We always trust our intuition because we are always aware, at some level, that we don't know what we don't know. So we lean on the part of us that has been quietly learning from everything, not just the things we can name.

AI does not have this. AI has no accumulated felt experience. It has patterns in text, not patterns in living.

How Our Brains Actually Work

When I think about how our brains retrieve information, it strikes me how different it is from how a large language model works.

Imagine you are working on a property management platform. A new, ambiguous, abstract problem lands on your desk. How do you solve it? Some people start from root causes. Some apply first principles thinking. Some go straight to references and analogies from past work. Everyone thinks with a different approach, and the reason they have different approaches is that each person is more comfortable with the method that has worked for them before, shaped by their own unique history of successes and failures.

This is why the same problem, given to five different product people, produces five different solutions. And the success criteria between them are different too. The explicit success metric might be the same: hit this revenue number, reach this engagement rate. But each person also carries implicit success criteria, tacit measures of quality that they hold themselves to. “This is the way I arrived at this solution, and I believe it will hit the business metric and satisfy something deeper about how I think good products should work.”

That “something deeper” is tacit knowledge in action. It is the part of the iceberg that is underwater.

The Tiny Pieces That Live in Your Brain

Even right now, as I write this, I am aware of how much I am leaving out. I have more context than I can fit into this post. I have more thoughts, more tiny pieces of information that connect to this topic in ways I can barely trace. And these pieces live in my brain in a way that is hard to explain. They don't sit in a neat index. They don't have tags or categories.

Sometimes a piece just pops up. Sometimes it surfaces when someone asks me a question I didn't expect. My fellows at Project X Vietnam ask me something during a mentoring session, and suddenly my brain activates some neuron, some connection, and pulls up a specific piece of knowledge that I had not thought about in months. A piece that I certainly never wrote down. A piece that I could not have retrieved on demand if you had asked me to list everything I know about the topic.

That is how tacit knowledge works. It is not organized. It is not searchable. It is activated by context, by proximity, by the specific contour of a question or a problem. And it is deeply, irreducibly personal.

An AI does not have this. An AI retrieves information through pattern matching over text. It does not have the experience of a question lighting up an unexpected corner of its memory. It does not remember in the way we remember, where remembering is an active, creative act that reconstructs meaning in real time, colored by who we are in the moment of recall.

Output Is Easy. Input and Process Are Human.

I want to be fair. AI is making excellent work on the output layer. I experience this myself every day.

When I draft an email, AI makes it detailed and clear. When I write a ticket description on Linear to hand over to my engineers and designers, the result is thorough, well-structured, precise. Sometimes I even have to trim the details just to make it feel more human, more direct, because the AI tends toward a kind of perfection in explicitness that can feel clinical.

So yes, the output is something AI does very well. Everything explicit, everything that is about articulating what has already been decided, AI has learned to do that to a remarkable degree.

But here is what AI cannot do.

The input: knowing what to build, what question to ask, what problem is actually worth solving. The process of creating the input, the messy, non-linear, deeply human process of thinking through ambiguity and arriving at a direction. That process is something that still depends entirely on us.

And the judgment: knowing which output is the good one and which one to throw away. Looking at three perfectly articulated options and knowing, in your gut and your experience, that the second one is right even though all three look equally reasonable on paper. That is also something only humans can do, because it requires the accumulated weight of every decision you have ever made, every failure you have absorbed, every intuition you have developed through living.

AI understands input and output. But it does not understand process. It does not understand how something became what it is. It sees the before and after but not the becoming. And the becoming is where all the tacit knowledge lives.

The Codification Trap

There is a natural instinct, especially in companies, to respond to all of this by trying to capture everything. Document all the processes. Write the runbooks. Record the meetings. Feed it all into the system.

But the act of writing knowledge down changes it. When you ask a master carpenter how they know when the wood is ready, they might say “you can feel it” or “it sounds different.” Force them to write a procedure, and you get a flat list of steps that misses the thing that actually mattered: the embodied judgment.

Tacit knowledge resists codification not because we are lazy about documenting, but because it is a fundamentally different kind of knowing. It is knowing how, not knowing that.

The Apprenticeship Gap

Tacit knowledge has always transferred through proximity. Through apprenticeship, mentorship, working alongside someone and absorbing how they make decisions.

A junior engineer does not learn system design from a wiki. They learn it by watching a senior engineer reject their pull request and hearing, in the hallway after, why that abstraction would have caused problems in six months. A product person does not learn prioritization from a framework. They learn it by sitting in a room with someone who has shipped ten products and watching how they react when the CEO asks for a feature.

This is exactly what I see at Project X Vietnam. The knowledge that transforms our fellows is not in the curriculum. It is in the mentoring conversations, the live projects, the moments when they make a decision and a more experienced person says “here is why that might not work” and then helps them see the shape of a problem they could not have seen on their own.

If AI handles more of the explicit layer, the first-line questions, the routine code, the summarization work, there is a risk we lose the context in which tacit knowledge used to transfer. The junior does not sit next to the senior because they can ask the AI instead. The apprenticeship moments were bundled with the explicit work. Remove the explicit work, and you might accidentally remove the transfer mechanism for the tacit.

That is the real danger. Not that AI replaces us, but that it removes the conditions under which we grow.

What I Actually Believe

So where does this leave us?

Not in a race against AI. That framing misunderstands the board entirely. AI is playing in the explicit layer: the codified, the written-down, the searchable. It is playing that game brilliantly. We will not beat it there, and we should not try.

Tacit knowledge is something we can create. We can nurture it. We can improve it along the way. Every time you build something and fail, you gain it. Every time you make a decision under uncertainty, you gain it. Every conversation that shifts your thinking by a few degrees adds to it. It accumulates. It compounds. And AI can never reach it, because it was never on any server, never in any dataset, never in any form that a model could ingest.

The human edge is the 80 to 95 percent that never got written down. It is the knowledge that transfers through presence, through doing, through the kind of understanding that only comes from living a life and paying attention.

The challenge is not to compete with AI. It is to get serious about the things AI has shown us it cannot touch. Invest in the relationships. Invest in the apprenticeships. Do the real work. Make the hard decisions. Document what you can, but know that the most valuable thing you carry will never fit in a document.

Because if I am an AI looking at a human, the most intimidating thing is not what they have written down. It is everything they have not. It is the vast majority that lives in their neurons, in their instincts, in the way they pause before answering a question because something does not feel right, even when they cannot say exactly what.

That is the human edge. And it is not going anywhere.