The Epistemology of Cognitive Uploading
Steven Johnson, Artificial Intelligence, and the Evolution of Augmented Cognition
A brief housekeeping note at the start. I am changing the publishing schedule for The Augmented Educator: new essays will now arrive every Thursday in early access for paid subscribers, then become free to everyone the following Tuesday.
My thanks to all of you for your continued interest and support. I truly appreciate it.
Michael G. Wagner (The Augmented Educator)
There are perhaps two ideas that shaped how I think about media and technology more than any others. One is Clayton Christensen’s account of disruptive innovation. The other is Steven Johnson’s “Sleeper Curve.” I came to the second in the early 2000s, when I was writing about computer games and trying to argue that the medium did real intellectual work, something many critics refused to acknowledge.
At that time, Johnson had already made that case, and he made it more persuasively than I could. That is why, when he speaks, I have learned to listen.
What has stayed with me from Johnson’s earlier work is not any single verdict about video games but a habit of mind. It is the recognition that our judgments about what is valuable and what counts as genuine thinking are far less stable than they feel from the inside. What looks self-evident in its moment—that comic books rot the brain or that video games cause real-world violence—has a way of looking outdated a generation later.
That instability is worth holding onto right now, because we are once again being told, with enormous confidence, exactly how a new technology is destroying our minds.
Johnson also happens to be the co-founder and editorial director of Google’s NotebookLM, so his interest in AI is that of a builder, not a bystander. His view comes down to a single phrase: cognitive uploading. The phrase inverts the anxiety that runs through nearly every conversation about AI, and that inversion is a far heavier lift than it first sounds.
The Sleeper Curve as a method
The Sleeper Curve was the organizing argument of Johnson’s 2005 book Everything Bad Is Good for You. The title is a wink at Woody Allen’s Sleeper, in which scientists of the future are baffled that twentieth-century humans never appreciated the health benefits of deep-fried food and cream pies.
Johnson’s claim was structurally similar and just as counterintuitive: the most disparaged products of mass culture had, by most of the measures we use to praise reading, been growing steadily more demanding.
Television had moved from tidy single-thread plots to the interwoven, ambiguous storylines of something like The Sopranos, which asks viewers to track a dozen relationships and simultaneously withhold easy moral judgment. And games, he argued, rewarded a constant cycle of probing a system’s hidden rules while continuously switching between immediate tasks and distant goals.
The crucial move was where Johnson located the cognitive value. It lay in the structure underneath, in the demand the medium placed on the user to build a working model of a complex system, not in the violence and spectacle that drew the moral panics. That distinction is the part of his thinking I have never stopped using.
It also translates directly to the argument about AI. The Sleeper Curve is not really a defense of television or games. It is a method. It tells you to stop asking whether a medium is good or bad and start asking what kind of thinking a given use of it requires. Applied to AI, asking whether the technology makes us smarter or dumber is too crude to be useful. What matters is what a given way of using it asks of us.
The case for cognitive decay
I need to acknowledge that the pessimistic case about the impact of technology on cognition is not without merit. It is old and serious, and there is actual evidence behind it.
It is old enough to predate the printing press. In Plato’s Phaedrus, Socrates famously warns that writing will implant forgetfulness in the souls of those who learn it, because they will cease to exercise their memory and rely instead on external marks. The complaint recurs at every technological threshold since, and it has rarely been completely wrong. Something is usually lost.
The contemporary version of this argument is more targeted. Critics such as Audrey Watters, the longtime ed-tech skeptic behind Hack Education, have spent years documenting how the industry recycles the same promises while delivering surveillance and standardization. And she is rightly wary of any story in which a Google product turns out to be the answer to any educational challenge.
The historian Jonathan Rees, writing about Johnson’s own use of AI to draft a book, similarly worried that feeding primary sources to a machine in order to generate a narrative is not a healthy substitute for the slow, irreplaceable labor of reading them yourself. The concern is that reading a stream of brief summaries is not the same activity as reading a book, and that as we offload more of the reading, we lose the intellectual stamina that deep reading was building.
There is experimental evidence for it as well. The psychologist Linda Henkel found, in a study at Fairfield University’s art museum, that visitors who photographed objects on a guided tour later recognized them less accurately than objects they had merely observed. Because the camera remembered for them, they paid the objects less attention. She called it the photo-taking impairment effect.
When we hand a task to an external system precisely so we can forget it, we do, in fact, forget it. This is the heart of what critics call the illusion of understanding: the product appears without the process, and the gap between them is invisible to the person holding the result. These findings are real. But the question is whether they describe the complete picture or only the most negative aspect of it.
That problem is, at its core, an epistemological one. How do we know when a tool has helped someone to really know something, rather than merely helped them produce something that looks like knowledge?
Uploading, not offloading
Johnson’s argument is that these findings describe only one use of AI, and that a second use, its structural opposite, has gone mostly unnoticed because we are too anxious to look for it.
The pessimistic case is really a case against the practice of cognitive offloading: treating the machine as an oracle, handing it a generic prompt, and accepting the output with no friction and no engagement. “Write me an essay on the causes of the First World War.” Here, the cognitive demand is exported wholesale. The user keeps the essay and loses the education. On that, Johnson and the critics agree completely. Offloading asks the machine to replace judgment. But uploading, Johnson argues, asks it to provoke judgment.
Cognitive uploading, in his account, runs in the other direction. Rather than offloading a task in order to forget it, the user uploads a curated body of trusted material, including primary sources or their own accumulated notes, into a system constrained to reason only over that corpus.
This is what NotebookLM is built to do: it is designed to answer from the sources in the notebook and to cite them. That does not make error impossible, but it changes what the reader can check — not just whether the output reads well, but where each claim is grounded. The AI becomes less a search engine than a connection engine, a place to test relations among sources rather than retrieve a single answer. It can hold tens of thousands of quotations in immediate recall, drawing links across them that a human memory would never surface.
Johnson’s own example describes this perfectly. Researching a book on the California Gold Rush, he uploaded a set of open-source histories and Indigenous accounts and began interrogating them. The system surfaced Maria Lebrado, granddaughter of the Yosemite chief Tenaya, and noted that she had returned to the valley near the end of her life, at almost ninety. From that thread, it proposed a narrative architecture: open with the old woman’s return, then flash back to the violence of her 1850s childhood.
Johnson recognized the shape immediately and compared it, half-laughing, to the structure of Titanic. The work he estimates would have taken weeks of archival synthesis arrived in minutes. The point here is not that the machine wrote the book; it did not, and could not. It just handed him a structure and left the writing and the judgment to him.
This is the real inversion. Uploading does not remove friction from the intellectual work. It relocates it. The friction moves away from finding and remembering every passage and toward the harder work of selection and judgment. Uploading only removes retrieval friction; it does not remove interpretive friction. In the best cases, it even increases it.
As Johnson puts it, the aim is to give a person more to think about, not less. The approach it demands is that of a gamer—probing what the system will do and actively steering toward a goal—rather than that of a passive consumer. Which is to say the Sleeper Curve has come back around, now aimed at the very technology most people assume is its refutation.
A note on terminology
I think we need to be precise about the term itself, because Johnson did not coin it, and the difference between his usage and the older one is worth clarifying.
“Cognitive uploading” has a prior life in the philosophy of mind. It belongs to the lineage of the extended mind, the thesis Andy Clark and David Chalmers advanced in 1998, that cognition is not limited to the brain. When we reliably recruit the environment to do cognitive work, the environment becomes part of the machinery of thought.
Working in that tradition, Axel Constant, Andy Clark, Michael Kirchhoff and Karl J. Friston gave “uploading” a specific technical sense, contrasting it with “offloading” in a way that does not map onto Johnson’s usage of the term at all. There, offloading is a temporary fix for one person’s task, while uploading means durable cognitive functions absorbed by the shared environment and outliving the individuals who made them.
Johnson’s uploading is a near-opposite in everything but spirit. Where the philosophers mean something collective and left in the world for others, he means something personal and computational: a private corpus, curated by one mind and animated by a model that talks back.
He has, in effect, borrowed a word for a species-wide phenomenon and pointed it at an intimate one. What survives the translation is the underlying claim of the extended-mind tradition: that thinking has never been confined to the brain, and that the right external structures genuinely enhance it.
What it looks like in a classroom
The distinction between cognitive offloading and uploading stops being abstract the moment it enters a school, because the same tool will reward either depending entirely on how it is used.
Take the most mundane use first. State standards and faculty handbooks are often dense to the point of being unnavigable. Uploading one to a grounded system and asking it for the exact learning objectives for a particular unit returns a sourced answer the user can check against the document. That is administrative friction lifted, cleanly.
The pedagogically interesting uses are at a higher level. A teacher can upload a university-level paper on photosynthesis and ask the system to re-explain it to a nine-year-old through the analogy of baking a cake. This generates differentiated material from a rigorous source while leaving the teacher responsible for checking where the analogy clarifies and where it distorts.
There is a simple test teachers can apply. A classroom use of AI is cognitively healthy when students still choose the sources, weigh the evidence, and test the model’s synthesis against their own understanding. It is cognitively corrosive when the model chooses the frame and supplies the language while the student does little more than approve the result.
Let’s look at three example assignments involving cognitive uploading.
In a source interrogation, students upload the assigned primary documents and ask the system where the accounts contradict, omit, or corroborate one another, and then argue with its answer. In an evidence audit, students take a claim the model produces and trace every part back to a cited passage, marking what it missed or overstated. And in a synthesis map, students let the model surface connections among their sources, but their grade rests on explaining which connections they accepted, which they rejected, and why.
In each of these examples, the machine does the retrieval and the student keeps the judgment, which is the whole of the matter.
A thinker worth following
I have spent the better part of two decades reading Steven Johnson, and the consistency of his arguments earns him my trust. From the defense of disreputable pop culture, through Where Good Ideas Come From and his histories of cholera and dynamite, to the design of NotebookLM, he has returned to one conviction: that the tools which look like they are dumbing us down are often the ones quietly asking more of us.
If this essay has done its work, it should send you to two places. The first is Johnson’s Substack, Adjacent Possible, where he is thinking through these questions in public and in real time, and where the cognitive uploading argument appears in fuller form than I can do justice to here.
The second is NotebookLM itself, if you have not yet tried it. The argument is more persuasive once you have uploaded a stack of sources you actually care about and watched the gaps in your own knowledge light up.
A final caution, in keeping with the skeptics, who are not wrong about the danger. The optimistic reading is not automatic. The same system that can be uploaded into can just as easily be offloaded onto. And in my experience, many users, left to their own devices, reach for the easy oracle. Our goal as augmented educators is to guide our students toward the correct path.
Uploading and offloading travel the same wires. The difference has never been the machine. It has always been us.
The images in this article are reinterpretations of scenes from the Woody Allen movie “Sleeper” and were generated with Nano Banana 2.
P.S. I believe transparency builds the trust that AI detection systems fail to enforce. That’s why I’ve published an ethics and AI disclosure statement, which outlines how I integrate AI tools into my intellectual work.