Popper's World 3: Why AI Can Only Use the Knowledge Your Company Has Written Down

Organizational knowledge for AI agents lives in Popper's World 3: the part you wrote down. Why AI acts only on externalized knowledge, and how to build it.

In 1967 a philosopher of science stood up in Amsterdam and proposed a way of dividing reality into three parts. He was not thinking about your company. He was thinking about mathematics, libraries, and the question of whether a scientific theory is still true when nobody is currently thinking it. Almost sixty years later his scheme turns out to be the cleanest available answer to a very practical 2026 problem: why the AI you bought does so much less than the demo promised.

The short version is that your organization has knowledge in two very different places, and the AI can reach only one of them. Karl Popper gave those two places names. The knowledge inside your people's heads he called World 2. The knowledge you have written down, drawn, codified, and stored outside any single head he called World 3. An AI agent has no access whatsoever to World 2. It works on World 3 or it works on nothing. Most companies have spent decades running almost entirely on World 2, and have just discovered, the expensive way, that this was a bet against the future.

This article is about that distinction and what to do with it. It is genuinely useful, not a metaphor we are stretching for effect, and there is a small scholarly literature that has been making the connection for twenty years. By the end you will have a precise vocabulary for the thing every AI vendor gestures at and none of them name: the reason the model is rarely the problem, and your written knowledge almost always is.

Karl Popper introduced the three worlds in a lecture titled "Epistemology Without a Knowing Subject," delivered in Amsterdam in 1967 and collected in his 1972 book Objective Knowledge: An Evolutionary Approach. His own definition is worth quoting exactly:

"We may distinguish the following three worlds or universes: first, the world of physical objects or of physical states; secondly, the world of states of consciousness, or of mental states, or perhaps of behavioural dispositions to act; and thirdly, the world of objective contents of thought, especially of scientific and poetic thoughts and of works of art."

World 1 is physical: stones, stars, bodies, buildings, and the hardware your code runs on. World 2 is mental and subjective: your perceptions, your intentions, the felt sense of knowing how to handle a difficult client, everything that happens inside a conscious mind. World 3 is the strange one. It is the world of the contents of thought once they have been externalized into something that exists independently of any particular mind: theories, arguments, problems, numbers, books, blueprints, laws, a documented process. World 3 is man-made, and Popper was insistent on this point, but once made it takes on a life of its own.

The claim that makes World 3 more than a filing cabinet is what Popper called its autonomy. He listed it among his supporting theses: "the third world is largely autonomous, even though we constantly act upon it and are acted upon by it." His favourite illustration was arithmetic. We invent the sequence of natural numbers, so the numbers are our product. But once they exist, they have consequences we did not design and could not have anticipated: "The series of natural numbers which we construct creates prime numbers, which we discover, and these in turn create problems of which we never dreamt." The primes were not in the plan. They were waiting inside the thing we built, and we found them by exploring it. Hold on to that image. It is exactly what a good AI agent does when it reads your documentation, and we will come back to it.

Popper was careful to separate his World 3 from Plato's realm of Forms, which the scheme superficially resembles. Plato's world was eternal, divine, and entirely true. Popper's is the opposite: "My third world is man-made and changing. It contains not only true theories but also false ones, and especially open problems, conjectures, and refutations." Your company's World 3 is the same. It is full of out-of-date docs, contradictory policies, and half-finished thinking. That is not a failure of the category. That is the category working as designed.

The cleanest demonstration of why World 3 matters is a thought experiment Popper offers near the end of the essay. Imagine all our machines and tools are destroyed, and all our skills with them, but the libraries survive and we keep the ability to learn from them. After much suffering, he says, civilization recovers. Now imagine the same catastrophe, except this time the libraries burn too. Recovery does not happen, or takes millennia, because the knowledge embodied in those books is gone and there is no longer anything to re-learn it from. The difference between the two scenarios is the entire value of World 3. It is knowledge that survives the death of the knower. Popper's phrase for it, the title of the lecture, was knowledge "without a knowing subject."

That phrase is the whole article. An AI agent is a knowing process without a knowing subject. It can only meet your knowledge in the place where knowledge already exists without a knower: World 3.

You do not need to adopt any philosophy to see the three worlds inside your own organization, because they are already there and you manage all three badly in different ways.

Your World 1 is the physical and infrastructural layer: the offices, the laptops, the servers, the production database as a set of bytes on a disk. Your World 2 is the collective mind of your team: what your head of sales has learned about which objections are real, the instinct your lead engineer has for where the system breaks under load, the thing your founder knows about why the second-largest customer almost churned in 2024. Your World 3 is everything you have externalized: the codebase, the wiki, the Slack archive, the recorded calls, the spreadsheets, the contracts, the runbooks, the README nobody updated.

The connection to information systems is not a stretch we are inventing. Joseph Firestone and Mark McElroy built an entire theory of knowledge management on exactly this mapping in their 2003 book Key Issues in the New Knowledge Management, arguing that physical information systems are World 1, subjective and tacit know-how is World 2, and the objective, codified knowledge of the organization is World 3. Their complaint, two decades ago, was that the field had become obsessed with capturing tacit knowledge and blind to the World 3 it already had. That complaint reads as prophecy now.

Here is the part most management writing gets wrong. Your organization's real, operational knowledge is not the sum of what your people know. It is much smaller than that, and it is a specific subset: the part that has made it into World 3. Everything else is a liability dressed as an asset, because it walks out the door at five o'clock and sometimes does not come back. The organizational-memory literature has a precise way of saying this. In their foundational 1991 paper "Organizational Memory," James Walsh and Gerardo Ungson describe knowledge as retained across several stores, of which individuals are only one, alongside culture, routines, structures, and the physical setting. The organization "remembers" things no single employee knows, and it forgets things every time someone leaves. World 3 is the part of that memory that does not depend on anyone in particular still being employed.

For most of business history, the gap between World 2 and World 3 was a manageable nuisance. You wrote down what you had to, kept the rest in people's heads, and accepted the occasional bus-factor disaster. That arrangement just stopped being viable, and it is worth being precise about why.

Running a company mostly out of World 2 has real advantages, which is why intelligent organizations did it on purpose. Tacit knowledge is fast, cheap to maintain, and self-updating. The colleague who knows how the deployment really works does not need to keep documentation in sync; her understanding updates itself every time she does the job. Michael Polanyi, in his 1966 book The Tacit Dimension, opened with the line that has anchored this whole field since: "we can know more than we can tell." A great deal of what makes an organization good is precisely this tellable-only-with-difficulty knowledge, and forcing all of it into documents would be slow, lossy, and demoralizing.

So companies developed a sensible division of labour. The knowledge that was expensive to externalize stayed in World 2, carried around inside people who could apply it on demand and re-derive the World 3 version whenever it was actually needed. A consultant could hold a methodology in her head and write the specific deck for the specific client. Harvard's Morten Hansen, Nitin Nohria, and Thomas Tierney described this in their 1999 article "What's Your Strategy for Managing Knowledge?" as the choice between codification, the people-to-documents approach, and personalization, the person-to-person approach. Their finding was that the best firms picked one as their dominant mode rather than splitting the difference. Plenty of excellent companies, including the consultancies they studied, chose personalization. They bet on World 2.

That bet had one load-bearing assumption: that the only thing which ever needed to read your knowledge was a human being, and humans come with World 2 attached. A new hire learns by sitting next to someone. A question gets answered in a hallway. The knowledge never had to fully leave anyone's head because there was always a head available to consult.

The assumption held for a century. It is now false. There is a new reader in the building, and it has no head.

An AI agent is a consumer of your knowledge that cannot do any of the things a human colleague does to compensate for thin documentation. It cannot sit next to your senior engineer for three months. It cannot read the room. It cannot infer from a teammate's tone that the official process is not the real process. It has no World 2 of its own about your specific company, and no way to access yours. It can read World 3, and it can read nothing else.

This is the structural fact underneath the entire enterprise-AI discourse, and it is usually described in narrower technical language. When practitioners in 2025 and 2026 started saying that "context engineering" was replacing "prompt engineering," they were noticing this. Gartner's framing of context engineering is that enterprise AI fails for lack of organizational context, not for lack of model intelligence. A more recent strand of academic work describes the work of preparing a company for agents as externalization for a machine consumer: the conversion of tacit organizational knowledge into structured, machine-readable representations. These are all the same observation. The model is a reasoning engine that arrives knowing everything on the public internet and nothing about you. The only channel through which it can learn about you is your World 3.

This is why the model rarely turns out to be the bottleneck. Swapping a frontier model for a slightly better frontier model changes very little if the thing you are pointing it at is a Notion graveyard and a Slack history. The binding constraint is the size, quality, and machine-readability of the knowledge you have externalized. A company with a thin World 3 and the best model on the market will lose to a company with a rich World 3 and a mediocre one. The leverage moved, and it moved to a place most organizations have systematically underinvested in for their entire existence.

There is, to be fair, one other writer who has made the Popper connection explicitly. Matthew Davis, in a February 2026 essay called "AI as World 3 Native," argues that AI is a World 3 engine running on World 3 fuel, and that the technology is reviving a category philosophers had let go dormant. He is right about the philosophy. What that piece does not do, and what matters if you are running a 12-person company rather than thinking about AGI, is connect the idea to the operational reality of where your knowledge currently sits and what it would cost to move it. That is the part worth your attention, because the evidence on what happens when companies skip it is now substantial.

The headline number of the year came from MIT's Project NANDA, whose August 2025 report The GenAI Divide: State of AI in Business 2025 found that roughly 95% of enterprise generative-AI pilots delivered no measurable business return. The study was not a survey of opinions; it combined 150 leader interviews, a survey of 350 employees, and an analysis of 300 public AI deployments, against a backdrop of 30 to 40 billion dollars of enterprise investment. The authors located the cause in what they called a learning gap: the systems did not retain feedback, adapt to context, or improve in contact with the actual work. Read through Popper's lens, "adapt to context" means "connect to the organization's World 3," and most of the failed pilots had very little World 3 to connect to.

The number that names the bottleneck most directly comes from McKinsey's 2025 State of AI research, which reported that eight in ten companies cite data limitations as a roadblock to scaling agentic AI. The same body of work found that while 88% of organizations now use AI in at least one function, only around a third have begun to scale it, and fewer than 10% have scaled agents to deliver tangible value. Gartner, looking forward, predicted in June 2025 that over 40% of agentic-AI projects will be cancelled by the end of 2027, citing unclear business value and inadequate context among the reasons.

It is worth being careful with these figures, because the vendor internet has already begun rounding them into whatever shape sells a product. The MIT, McKinsey, and Gartner numbers above are from named institutions with stated methods, and they point the same direction: the failures cluster not around model capability but around the organization's readiness to feed a model anything useful about itself. "Data limitations" and "lack of context" are management-consulting words for a thin, messy, machine-illegible World 3.

None of this should be read as a reason to wait. It is a reason to do the unglamorous thing first. The companies in the successful 5% did not have better models. They had a body of externalized knowledge worth pointing a model at, and a clear, narrow task that lived inside that knowledge.

If the work of becoming useful to AI is the work of growing your World 3, then the relevant question is how knowledge gets from World 2 into World 3 in the first place. Management theory has a precise and well-tested answer, and it predates the current moment by thirty years.

In The Knowledge-Creating Company (1995), Ikujiro Nonaka and Hirotaka Takeuchi described knowledge creation as a spiral through four modes, known by the acronym SECI. Two of them concern us. Externalization is the conversion of tacit knowledge into explicit knowledge: articulating what was previously only known in the doing, through dialogue, metaphor, analogy, and worked examples. Combination is the conversion of explicit knowledge into more explicit knowledge: taking documents that already exist and recombining them into something more systematic. Nonaka considered externalization the pivotal and hardest mode, the one where genuinely new shareable knowledge is born. He was describing it as a route to human learning and innovation. It turns out to be, line for line, the description of preparing your company for an AI agent. Externalization grows your World 3 from your World 2. Combination is most of what the agent itself will then do.

This reframes the project in a way that is both humbler and more actionable than "adopt AI." You are not adopting a technology. You are running an externalization program, and the AI is simply the most demanding consumer your World 3 has ever had, the one that will not paper over the gaps with hallway conversation. The discipline this requires is the discipline of writing things down well: not everything, not as a bureaucratic exercise, but deliberately, where it counts, in a form a machine can read losslessly.

The form matters as much as the act. A 2026-grade World 3 is not a pile of PDFs and a Notion workspace behind an OAuth wall. It is knowledge an agent can reach through a stable, low-friction interface, in atomic enough units that it can retrieve the relevant paragraph rather than the whole forty-page strategy doc, with a version history that says who changed what and why. This is why so many AI-native teams are quietly moving their ground-truth knowledge into git-versioned markdown that humans and machines read equally well. The substrate that is good for a machine turns out, almost by accident, to be good for the humans who are actually paying attention. We have written elsewhere about why Notion is the floor and not the ceiling for this; the World 3 frame is the reason the argument generalizes beyond any one tool.

Return to the prime numbers. Popper's deepest claim about World 3 was not that it stores knowledge but that it generates it: once you externalize something, it carries consequences you did not put there, and those consequences can be discovered. The natural numbers were ours; the primes were waiting inside them. World 3 is autonomous in the sense that it contains more than its makers consciously intended.

This is the part of the philosophy that pays a dividend, because it is precisely what a capable agent does when it crawls a rich, well-formed World 3. It does not just retrieve what you wrote. It finds the implications of what you wrote: the contradiction between the pricing doc and the contract template, the customer segment that three separate reports describe without anyone having named it, the runbook step that the incident history shows always fails. These were latent in your World 3 already, true before anyone noticed them, in exactly Popper's sense that an unsolved problem exists objectively before it is discovered. The agent is a machine for exploring the autonomous consequences of your own externalized knowledge.

This is also why the investment compounds in a way that World 2 never could. Knowledge held in heads does not cross-reference itself while you sleep. Knowledge in a well-structured World 3 becomes a space that can be searched, recombined, and mined for the problems it implies, and the better organized it is, the more of those latent problems become reachable. A thin World 3 has almost no autonomous structure to discover. A rich one has primes hiding in it. The return on externalization is not linear in the number of documents; it is closer to combinatorial in the connections between them, which is the whole reason knowledge graphs and retrieval systems became the center of gravity of enterprise AI in 2026.

It also sharpens what you should be afraid of. An agent exploring your World 3 will surface your contradictions and your rot as readily as your insight, because it cannot tell the difference between a current policy and an abandoned one if you never marked which is which. The autonomy of World 3 is indifferent to your intentions. That is an argument for curation, not for staying out of the water.

The abstract version is settled, so here is the concrete one. For a team of five to twenty people, the World 3 worth building deliberately, and the form that makes it legible to both colleagues and agents, looks roughly like this.

• Decisions and their reasons. Not just what you chose but why, and what you rejected. A decision without its reasoning is the World 3 equivalent of a number with no operations defined on it: present but inert. This is the single highest-value externalization most small companies are missing.
• Processes as they actually run. The real deployment process, the real way a deal moves through the pipeline, including the steps that are technically against policy but happen anyway. An agent that learns the official fiction will confidently do the wrong thing.
• The customer and domain model. The vocabulary your company uses, what each term means, which customer types exist and how they differ. This is the layer that lets an agent's general intelligence become specifically about you.
• Failures and incidents. What broke, what the fix was, what the underlying cause turned out to be. This is the highest-signal training material for any agent doing operational work, and it is almost always trapped in a closed Slack thread, which is World 2 with extra steps.
• The contracts, policies, and constraints. The things that are true whether or not anyone remembers them, the GDPR boundary, the data-handling rule, the commitment made to a customer in 2024. These are the parts of World 3 where being machine-readable is not a convenience but a safety requirement.

The unifying test is Popper's library experiment, scaled down to your company. If your two most senior people left on the same Friday, how much of what makes your company good would survive the weekend? Whatever the honest answer is, that is the current size of your World 3. Everything else is World 2 you have been lucky enough to keep employed.

There is a useful intermediate concept here from the sociology of knowledge. Susan Leigh Star and James Griesemer's 1989 idea of "boundary objects," artifacts that are robust enough to mean something across different groups while staying flexible enough for each group to use locally, describes what good World 3 documents do inside an organization. A well-written domain model is a boundary object between sales and engineering. Increasingly it is also a boundary object between your people and your agents, which makes the quality of that shared artifact matter more than it ever has.

The World 3 frame is powerful, which means it is also possible to overapply it, and the honest version of this argument has to mark where it stops.

The first limit is the one Polanyi insisted on and the management theorist Haridimos Tsoukas pressed hardest in his 2003 essay "Do We Really Understand Tacit Knowledge?" Some knowledge is not in World 2 because nobody got around to writing it down. It is in World 2 because it is, in its nature, not fully tellable. The feel for a negotiation, the judgment about whether a candidate is right for the team, the craft sense of when a system is about to misbehave, these resist externalization, and the attempt to force them into documents produces a confident, lossy, and sometimes dangerous facsimile. Tsoukas's warning is that tacit and explicit knowledge are not two ends of a continuum you can slide along at will; they are two sides of the same coin. A program that tries to externalize everything will waste enormous effort on the parts that cannot move and will produce documents that are worse than the silence they replaced. The skill is in choosing what genuinely can and should become World 3, and leaving the rest in the heads that hold it well.

The second limit is that externalized knowledge can be wrong, and a confident wrong document is more dangerous than an absent one, because both a human and an agent will trust it. Growing your World 3 without curating it is a way of scaling your contradictions. The work is not "write more down." It is "write the right things down, mark what is current, and retire what is not."

The third limit is the oldest one in the field. Jeffrey Pfeffer and Robert Sutton called it the knowing-doing gap in their 2000 book of that name: organizations routinely know what to do and fail to do it, and a richer knowledge base does not close that gap on its own. World 3 is necessary for AI to help you. It is not sufficient for the help to translate into changed behaviour. If your problem is that people do not act on what the company already knows, more documentation and a smarter agent will not fix it, and a consultant who tells you otherwise is selling.

The frame earns its keep when your real constraint is that an AI, or a new hire, or a stretched team, cannot reach knowledge that genuinely exists somewhere in the building. That is a remarkably common situation. It is just not the only one.

Is "World 3" just a fancy word for documentation? It is broader and more precise. Documentation is one kind of World 3 artifact, but so is your codebase, your data model, your contracts, and your recorded incident history. The value of the term is that it draws the line in the right place: not between formal and informal, but between knowledge that exists outside any single mind and knowledge that does not. That line is exactly the line an AI agent can and cannot cross.

Why can't AI just learn the tacit knowledge by working alongside us? Because it has no persistent World 2 of its own about your company and no senses to absorb the tacit layer the way a human apprentice does. Current agents learn your specifics through what you externalize, in context, at the time you ask. There is active research on giving agents better long-term memory, but even that memory is itself a form of externalized World 3. The agent does not acquire tacit knowledge; it consumes explicit knowledge, which is why the explicit layer is the constraint.

Does this mean we should write down everything before adopting AI? No, and trying to is a classic failure mode. Most small teams should start with a narrow, real task and the slice of World 3 it depends on, rather than a boil-the-ocean documentation project. Externalize where there is a specific agent or a specific decision that needs it. The MIT finding that 95% of pilots fail is partly a finding about teams that either skipped the knowledge work entirely or tried to do all of it at once.

Where should our World 3 actually live? In something an agent can read losslessly through a stable interface, in small units, with version history. In practice that increasingly means git-versioned markdown for the ground-truth layer, with your existing wiki kept as a human inbox on top. The principle matters more than the tool: full content in one operation, atomic retrieval, a diffable history of who changed what and why.

Is the Popper connection actually established, or are you stretching it? It is established. Joseph Firestone and Mark McElroy built a knowledge-management theory directly on Popper's three worlds in 2003, and there is a small information-systems literature mapping organizational knowledge onto the same ontology. The 2026 contribution is not the mapping; it is that AI agents make the World 3 layer economically critical in a way it never was when the only reader was a human with a World 2 attached.

The useful thing about the World 3 frame is that it converts a vague anxiety, "we should be doing something with AI," into a concrete and answerable question: what does this company know, where does that knowledge currently live, and which parts of it could a machine reach if we externalized them well? That question has a real answer for your specific organization, and finding it is most of the work.

It is also, more or less exactly, the conversation a Garden audit is built around. We map where your knowledge actually sits, World 2 against World 3, mark the gap that is keeping AI from being useful to you, and build a small, real system against the slice that matters rather than a slide deck about the whole. We run our own agentic systems in production, on our own externalized knowledge, so this is a description of how we work and not a theory we are trying out on you.

If you are trying to figure out which of your knowledge belongs in World 3 and what it would take to put it there well, that is the conversation we have. Email a@gardenresearch.eu.