The workflow has become familiar enough to feel inevitable. You encounter a paper, an article, a transcript, or a book. Rather than reading it slowly and taking notes by hand, you paste the text into an AI assistant and ask it to extract the key ideas, identify the central argument, and produce a structured summary formatted for your note-taking system. Within thirty seconds, a clean, well-organised document exists in your knowledge base. The source has been processed. The folder count has increased by one. The system has grown.
What has not grown, in most cases, is you.
This is not an argument that AI-assisted note-taking is without value. It is an argument that the value it provides is narrower than it appears, that it is often mistaken for something it is not, and that the mistake has consequences that compound over time. The consequence of building a knowledge management system around notes you did not earn is not immediately visible. It becomes visible the first time you need to use what you think you know — in an argument, in a design decision, in a moment of original thinking — and discover that the archive is full but the understanding is absent.
The Promise and the Confusion
Personal Knowledge Management as a discipline emerged from a genuine insight: that knowledge work, done seriously over a long period, produces more intellectual raw material than any unaided memory can organise and retrieve. The filing cabinet, the commonplace book, the index card system, and the digital note-taking application are all responses to the same problem — how do you preserve the insights you encounter, connect them to one another, and make them available when they are relevant?
The insight underlying PKM, when articulated by its most serious practitioners, was never that the system would do your thinking. It was that the system would serve your thinking — that by externalising the retrieval burden, you could free cognitive resources for synthesis, connection, and original work. The notes were meant to be nodes in a network of understanding that existed, primarily, in your mind. The system was a scaffold, not a substitute.
What AI tools have done, at their most aggressive, is invert this relationship. Instead of the system serving the mind, the system has begun replacing it. Instead of notes that record genuine encounters with ideas — the marginal annotations, the reformulations in your own words, the questions you could not yet answer — the system now contains summaries that were produced by a machine on your behalf, without your cognitive engagement, and filed under headings that suggest comprehension you may not actually possess.
The archive looks the same. The understanding is a different matter entirely.
What Encoding Actually Requires
To understand why this distinction is not trivial, it helps to understand something about how knowledge is actually acquired and retained.
The cognitive science literature on this subject is unusually consistent for a field in which controversy is the default. The process by which information moves from transient working memory into durable long-term memory is not passive. It requires active engagement with the material: retrieving it under conditions of difficulty, reformulating it in your own words, connecting it to things you already know, and applying it to problems that require you to adapt it. These processes are not peripheral to learning. They are the mechanism of it.
The specific name for the most important of these mechanisms is elaborative encoding. When you encounter a new idea and work to connect it to your existing knowledge — asking what it reminds you of, what it contradicts, what it would imply if true, what you would need to believe in order to accept it — you are creating a richer, more connected memory trace. The idea does not sit in isolation. It is woven into the existing structure of what you know, which means it can be retrieved by many different routes and applied in many different contexts.
AI-generated notes, by their nature, bypass elaborative encoding almost entirely. The cognitive work that encoding requires — the reformulation, the connection-making, the confrontation with difficulty — was performed by the AI, not by you. The output exists as text in a file. Whether it exists as understanding in your mind depends entirely on what you do with that text after it has been produced — and the design logic of most note-taking workflows treats filing as the final step, not the first.
The irony is precise: the part of the note-taking process that produces understanding is not the recording. It is the processing that precedes it. AI tools have made the recording frictionless and left the processing — the only part that matters cognitively — entirely to chance.
The False Archive
There is a particular kind of confidence that a well-stocked knowledge base produces, and it is worth examining carefully, because it is not always earned.
When you have a folder titled "Epistemology" with forty-seven notes inside it, cross-referenced and tagged and connected by bidirectional links, you are entitled to feel that you have engaged seriously with the subject. And if you wrote each of those notes yourself — if each one represents a genuine encounter with an idea, a reformulation in your own language, a connection you made to something you already understood — then the folder is a reasonable proxy for your engagement with the subject.
But if forty-three of those forty-seven notes were produced by an AI summarising sources you gave it, the folder is not a knowledge base. It is a library — a collection of other people's formulations of ideas, organised by a system that you maintain but did not intellectually inhabit. The difference between a library and a knowledge base is not the number of entries. It is whether the material has passed through a mind.
This matters practically. When you need to apply what the folder contains — when you are writing an argument that requires epistemic reasoning, or evaluating a claim about how people come to know things, or trying to understand why a particular philosophical position seems compelling but wrong — you will reach into the folder and find well-organised text that does not respond to your needs with the flexibility that genuine understanding would provide. You can copy from it. You can paraphrase it. But you cannot think with it, because thinking requires a kind of internalized structural knowledge that the folder never helped you build.
The archive is full. The understanding was never acquired. These two things can coexist indefinitely, which is what makes the situation so easy to overlook.
The Compounding Problem
If the consequence of AI-generated notes were simply that you knew less than you thought, it would be a manageable problem. You could discover the gap, do the actual intellectual work, and close it. But the consequence is somewhat worse than that, because knowledge structures are not flat. They are hierarchical. What you understand at one level determines what you can understand at the next.
Consider the situation of someone who has accumulated, through AI summarisation, a well-organised set of notes on a technical domain — say, the epistemology of scientific knowledge, or the history of a particular philosophical debate. The notes are accurate as far as they go. They correctly identify the key positions, the main arguments, the important figures. But the person who accumulated them did not do the cognitive work of building genuine understanding at the foundational level.
When that person attempts to advance — to engage with secondary literature, to form original views, to read arguments that assume and build on the foundational material — they will encounter a problem that is not obviously traceable to its source. The advanced material will feel harder than it should. Connections that should be obvious will not be obvious. Claims that should be immediately evaluable as consistent or inconsistent with the foundational positions will require re-checking, because the foundation was never truly internalised.
They will, in many cases, return to the AI and ask it to summarise the advanced material as well. The archive grows. The gap between what the archive contains and what the person understands does not narrow — it widens, because each layer of AI-generated notes adds to a structure that has no genuine cognitive foundation. You cannot build reliably on ground you never broke yourself.
This is the compounding problem. Each generation of note-taking that bypasses genuine understanding creates a dependency on further bypass. The strategy that saves time in the short term has a long-term cost that is difficult to perceive until it has become severe.
What Zettelkasten Was Actually Trying to Do
It is instructive to return to the practice that inspired the modern PKM movement, because its purpose is frequently misunderstood in ways that are directly relevant here.
Niklas Luhmann's Zettelkasten — the slip-box system he maintained for decades and from which he produced an extraordinary volume of original sociological work — was not, at its core, a retrieval system. It was a thinking system. The act of writing a note, in Luhmann's practice, was inseparable from the act of understanding what you were writing about. You could not add a card to the box without formulating an idea in your own words, identifying where it connected to existing cards, and making the connection explicit. The system forced genuine engagement. It was, deliberately, slow.
Luhmann himself described the process as a conversation — between the card he was writing and the cards already in the box, between the idea he was currently working with and the accumulated record of his prior thinking. The value of the system was not the cards themselves. It was the thinking that writing and filing them required, and the unexpected connections that emerged when you were forced to confront an idea in the context of everything else you had ever written about.
A Zettelkasten populated by AI-generated summaries is not a Zettelkasten. It is a database. Both have entries. Only one generates thinking. The distinction is not sentimental. It is functional. The system worked because of the cognitive friction it imposed, not in spite of it.
This is the same point that applies to every serious note-taking methodology, from the commonplace book tradition to the progressive summarisation techniques advocated by contemporary PKM practitioners. The methodology is the discipline of forcing your own mind to engage with the material. Remove the forcing function, and what remains is an organised repository of other people's thinking. Useful, perhaps, as a starting point. Genuinely yours in no meaningful sense.
The Stoic Distinction That Applies Here
There is a Stoic concept that maps onto this problem with unusual precision, and it is one that does not always receive the attention it deserves in discussions of knowledge management.
Epictetus distinguishes repeatedly between things that are eph' hēmin — up to us, within our power, genuinely ours — and things that are not. The distinction is usually discussed in the context of emotional equanimity: what can I control, and what must I accept? But it has an epistemological application that is equally important.
The understanding you have built through your own intellectual effort is eph' hēmin in the most complete sense. It is the product of your reasoning, your attention, your capacity to connect and evaluate and revise. It can be applied flexibly, adapted to new situations, defended under pressure, and built upon without limit. It belongs to you in the way that a skill belongs to you — not as an object you possess, but as a capacity you have developed.
The content of a note you did not write belongs to you in none of these senses. You can retrieve it. You can cite it. You can present it as your view if you are willing to misrepresent your relationship to it. But it does not function as understanding functions. It cannot be mobilised in the way that genuine knowledge can be mobilised. It is, in the Stoic terms, an external thing — dependent on the system, on the file, on the ability to look it up. Take the archive away and the knowledge disappears with it.
The Stoic insight is not that external things are worthless. It is that they are unreliable in a specific way — that they can be taken from you, or fail you at the moment you need them, in a way that genuine internal capability cannot. A knowledge base full of AI-generated notes fails you precisely when you need it most: when you need to think, not retrieve. When the problem is novel. When the argument requires flexibility. When the interviewer has asked the follow-up question that the summary did not anticipate.
What a Better Practice Looks Like
None of this implies that AI tools have no role in serious intellectual work. The question is where in the process they belong and what they should be asked to do.
There are tasks for which AI summarisation is genuinely valuable and cognitively appropriate. Getting a quick overview of a domain you have no prior exposure to, before you decide whether to invest serious time in it. Identifying which parts of a long document are relevant to a specific question you already understand. Generating a list of follow-up questions about a topic you have already thought about carefully. In each of these cases, the AI is doing preliminary work — work that precedes genuine intellectual engagement rather than substituting for it.
The tasks for which AI summarisation is least appropriate are precisely the ones it is most commonly used for: forming the core record of your engagement with an idea, building the foundational layer of your knowledge in a new domain, producing the notes that you will later build arguments from. These are the tasks that require your cognitive engagement most urgently, and they are the tasks from which the smooth efficiency of AI generation most readily releases you.
A better practice begins with the recognition that the note is not the product of reading. It is the evidence that reading produced understanding. If you cannot write the note in your own words, without reference to the source, you have not yet understood the idea. The note is the test, not the record. AI cannot take that test for you.
This does not mean that AI-generated summaries have no place in the workflow. They can serve as a prompt — a starting point for the harder work of reformulation, connection-making, and genuine integration. But they should come before your notes, not instead of them. The question to ask, every time you are tempted to file an AI-generated summary, is whether you could produce that summary yourself. If the answer is no, you have not yet done the work. The AI has done it for you, and the illusion that your knowledge base has grown is precisely that — an illusion.
The Archive Is Not the Point
Personal knowledge management, at its best, is a discipline of intellectual humility as much as intellectual organisation. It is an acknowledgement that minds are limited, that memory is fallible, and that the work of thinking well is continuous and requires external support. These are true and important things.
But the support that serious knowledge work requires is not the elimination of cognitive effort. It is the scaffolding that makes effort more productive — the system that connects ideas you have already internalized, surfaces relevant prior thinking when you need it, and reveals patterns across a body of genuine understanding that is too large to hold in working memory all at once.
An archive populated by AI-generated notes does not provide this support. It provides the appearance of it, which is considerably more dangerous than simply having no system at all. If you know you have not organised your thinking, you are motivated to do the work. If you believe your thinking is organised because your file system is full, you will not discover the deficit until you need the knowledge and find the shelf empty.
The goal of a knowledge management system is not to grow the archive. It is to grow the mind that the archive serves. These are not the same project. Only one of them is worth building.