Please do NOT feed the machine …
What Plato, Forster and McLuhan can teach us about AI, cognition and the future of knowledge work
We recently reviewed the material on our Substack as part of the redesign of the PolymathMind.ai website, and in doing so identified four core themes across the material. The first covered judgement and cognition under AI. The second examined labor-market disruption and the possible breaking of the master-apprentice model of knowledge work. The third looked at collective intelligence and the challenge of building organizations where humans and AI actually think better together. This fourth and final theme takes a step back to consider AI through a cultural and philosophical lens
The most useful frame for thinking about AI is probably not the engineering, well at least not for most of us. Nor is it this week’s product announcement, this quarter’s productivity data, or the latest claim about which model now beats which benchmark. These things matter, but they are not enough. The more durable frame is the history of how humans have negotiated with technologies that participate in cognition.
AI is often discussed as if it were primarily a capability question. What can the model do? Which tasks can it automate? How good is the output? Where does it outperform humans? These are useful questions, but the deeper question is what happens when a technology becomes part of the way humans think, know, judge and act.
A lot of the AI conversation is understandably dominated by two streams. The first is technical: capabilities, agents and workflows. The second is topical: this week’s announcement, the latest start-up, the newest regulatory consideration, what have Sam Altman or Dario Amodei just said?. Both streams have their place, but neither is sufficient for senior leaders making decisions whose consequences will outlast any specific tool.
So what can a longer cultural and philosophical frame tell us. It helps us see AI as the latest chapter in a much older story: the story of cognitive delegation. Humans have always built tools that extend, store, reorganize or redirect our thinking. Clark and Chalmers’ extended mind thesis made this point in modern philosophical terms: cognition does not always stop at the boundary of the skull. Its not just biological. It can be distributed across people, artefacts, tools and environments. Risko and Gilbert later described cognitive offloading as the use of external tools to reduce internal cognitive load.
So AI is not the first technology to support thinking. But it may be the first to participate in it … to mediate so much of professional knowing at such speed and scale.
In 1909, long before computers existed, E. M. Forster wrote The Machine Stops. Humans live isolated in underground rooms while a vast technological system provides everything: communication, knowledge, entertainment, social interaction. They no longer understand how the system works. They simply trust it. They deify it. When the machine eventually stops, they have lost both the knowledge and the capacity to fix it. They perish.
Forster was not writing about computers per se. He was writing about delegation — what happens when humans reorganize life around a system while allowing the underlying capacities of human agency to decay. Exploring Forster in 2026 feels very on-point in a way it perhaps did not in 1990. The story is not really about machines becoming powerful. It is about humans becoming dependent.
In Plato’s Phaedrus, Socrates lamented the invention of writing, fearing it would weaken memory and create the appearance of wisdom without its substance. He was not entirely wrong. Writing did change the architecture of human cognition. It allowed people to outsource memory, preserve argument, transmit ideas across time and construct traditions beyond the limits of individual recall. But it also became the foundation of every subsequent intellectual achievement. Philosophy, law, science, history and theology all depended on the thing Socrates feared. The interesting question was never whether to adopt writing. It was how to adopt it without losing what came before.
That question returns with AI, but with higher stakes. Writing allowed us to offload memory. Search allowed us to offload retrieval. Sparrow, Liu and Wegner’s work on the “Google effect” showed that easy access to information changes what people remember and how they approach recall. Calculators and spreadsheets allowed us to offload calculation. Generative AI allows us to offload parts of interpretation, synthesis, argument, judgement formation and meaning-making. That is a step change.
AI does not merely help us remember or retrieve. It proposes what matters. It summarizes the evidence. It drafts the conclusion. It suggests the next step. It often supplies the first version of the answer. The user is no longer simply working with data, information or evidence. They are working with an AI-mediated synthesis of that material.
This is why the word “tool” is beginning to feel inadequate. A hammer extends the hand. A spreadsheet extends calculation. But generative AI increasingly sits between the professional and the material they are trying to understand. It becomes an intermediary between humans, tasks and data. That is a different kind of relationship.
The mid-twentieth century cybernetics tradition saw this more clearly than much of today’s AI commentary. Norbert Wiener, Joseph Weizenbaum and Marshall McLuhan were not primarily worried about machines becoming intelligent in the narrow sense. They were worried about humans reorganizing thinking around machines. Wiener observed that automatic machines could displace cognitive labor as the steam engine had displaced physical labor. Weizenbaum, who built one of the first chatbots in the 1960s, was disturbed by how readily users attributed understanding to a system that had none. McLuhan argued that technologies do more than extend human capabilities. They reshape the structure of perception and thought. His famous line that “the medium is the message” has become so familiar that we often gloss over it. But a medium does not simply carry content … it changes the conditions under which content is perceived. AI may not just be a productivity tool. It may be a new cognitive medium.
Why on earth does that matter? It matters because knowledge is not simply an asset. It is tempting, particularly in organizations, to talk about knowledge as if it were a thing: something to be captured, stored, transferred, retrieved or generated. But in real work, knowledge is inseparable from knowing. Knowledge is not just encoded information. It is active, contextual, social and directed towards action. People know by probing, testing, comparing, arguing, noticing, remembering, interpreting and acting (Blackler, 1995)
This is where AI becomes genuinely consequential. Generative AI does not merely provide additional information to professionals. It can change the process through which knowing is accomplished. It can decide what is surfaced and what is hidden. It can turn messy material into a fluent synthesis. It can smooth over disagreement. It can impose structure before the human has struggled with the ambiguity. This may be useful. It may also be dangerous.
Alvarado describes AI as an epistemic technology: a technology involved in the production, organization and justification of knowledge. That seems right. AI is not simply doing work. It is shaping what information is available, what interpretations seem plausible, what claims appear credible, and what answers arrive first.
This is the point at which the philosophical question becomes practical. If a technology mediates what people know, then it also mediates what they notice, what they ignore, what they trust and what they feel responsible for. The risk is not only that AI might produce a wrong answer. The risk is that it might quietly reorganize the route by which people arrive at an answer. That is a very different problem from the one implied by the phrase “AI adoption”. A tool is introduced. Users are trained. Workflows are redesigned. Productivity is measured etc. etc.. The organization moves forward. But mediation is messier. It asks what now stands between the human and the work. It asks what kind of contact with reality is being preserved and what kind is being replaced by synthesis. It asks whether people are still participating in knowing, or mainly consuming outputs that have already shaped the field of possible thought.
This is not an argument against AI. That would be both futile and mistaken. The history of cognitive technologies tells us that resistance to the technology itself is rarely the wise move. Writing did not destroy thought. Printing did not destroy learning. The internet did not destroy intelligence, although it certainly reorganized attention. The real issue is not adoption versus rejection, but rather it is the ‘quality’ of adoption. Do we use AI in ways that extend human judgement, or in ways that allow the underlying human capacities to weaken because they are no longer practiced.
Plato’s anxiety about writing matters because it was never really just about writing. It was about what happens when a human capacity moves outside the person. Forster’s anxiety is relevant because it was never really just about machines. It was about what happens when a system becomes so convenient, so total and so trusted that people lose the ability to stand apart from it. Weizenbaum’s anxiety matters because it was never really just about chatbots. It was about how quickly humans attribute understanding to systems that manipulate language convincingly.
What are the right questions to ask … Who has authority? What counts as knowing? What kind of trust is appropriate? Where does responsibility sit? What should humans continue to practice? What should be delegated? What should never be delegated fully? What forms of direct contact with reality must be preserved? These are not abstract questions. They are leadership questions. They are capability-building questions.
AI does not merely stand outside the work. Increasingly, it participates in the work of knowing. It helps frame the problem, sort the evidence, generate the language, simulate the argument and produce the answer. In doing so, it changes the relationship between the professional and their own cognitive process.
The implications
Read more widely than the AI commentary itself. The most useful insights about AI are rarely written by AI commentators. Plato, Forster, Wiener, Weizenbaum and McLuhan saw the core dynamics earlier and often more clearly.
Distinguish capability from mediation. AI is not just a tool that does things. It is increasingly a medium that shapes how people encounter information, form interpretations and decide what deserves trust.
Protect epistemic friction. Leaders should remove wasteful friction, but preserve the productive friction created by challenge, ambiguity, dissent and uncertainty. Judgement needs resistance.
Design for engaged augmentation. The presence of a human in the workflow is not enough. Organizations need practices that require people to interrogate AI outputs, test assumptions, check omissions and integrate machine-generated answers with human experience and context.
Treat knowing as a practice, not a stock of information. Organizations that treat knowledge as content to be generated, stored and retrieved will miss the deeper issue. Knowing is something people do. AI can support that practice, but it can also bypass it.
Ask what contact with reality is being lost. When people work from summaries, synthetic outputs and machine-shaped first drafts, they may gain speed but lose texture. The question is not whether the synthesis is useful. It is whether enough direct engagement remains for judgement to stay grounded.
Take the philosophical questions seriously, especially when they look impractical. Questions about authority, trust, responsibility, meaning and judgement will determine more about AI deployment than many of the engineering questions will.
This is the work PolymathMind does — helping senior leaders close the gap between AI adoption and the human operating model around it. If your organization is wrestling with what to keep meaningfully human, how to protect the routes through which expertise is built, or how to make accountability real rather than decorative, those are conversations we’d welcome.
Where to go deeper
Read more on this subject in earlier PolymathMind Substack posts
Beware the Silicon Idols (early 2026) — Dune, Wiener, Forster, McLuhan — the literary and historical view
The Philosopher Kings of Silicon (early 2026) — Why AI labs are hiring philosophers — and what it means
Magic Intelligence in the Sky (13 Nov 2023) — Sam Altman’s ‘magic intelligence’ framing and what is being elided
Mostly Harmless (7 Sept 2025) — What Douglas Adams can teach us about AI
Bibliography
Alvarado, R., 2023. AI as an epistemic technology. Science and Engineering Ethics, 29, Article 32.
Blackler, F. 1995. Knowledge, knowledge work and organizations: An overview and interpretation. Organization Studies, 16(6), 1021–1046.
Clark, A. and Chalmers, D., 1998. The extended mind. Analysis, 58(1), pp.7–19.
Lebovitz, S., Lifshitz-Assaf, H. and Levina, N., 2022. To engage or not to engage with AI for critical judgments: how professionals deal with opacity when using AI for medical diagnosis. Organization Science, 33(1), pp.126–148.
Ramsoomair, N., 2025. Pressing matters: how AI irons out epistemic friction and smooths over diversity. Atlantis, 46, pp.42–55.
Risko, E.F. and Gilbert, S.J., 2016. Cognitive offloading. Trends in Cognitive Sciences, 20(9), pp.676–688.
Sparrow, B., Liu, J. and Wegner, D.M., 2011. Google effects on memory: cognitive consequences of having information at our fingertips. Science, 333(6043), pp.776–778.