The Extended Mind: You Have Always Thought with Tools

That false binary — AI makes you smarter versus AI makes you dumber — turns out to rest on an assumption so old it feels invisible. The assumption is that your mind lives inside your skull and nowhere else. Strip that assumption away, and the entire conversation about AI changes shape.

Here's the reframe that makes this section worth your time: the question of whether AI is a legitimate cognitive tool, whether it counts as "real" thinking, whether using it is somehow cheating — those questions can only be answered by first understanding what human cognition actually is and where it actually happens. And the answer, backed by decades of philosophy and cognitive science, is considerably stranger and more interesting than the folk theory most people carry around.

The territory ahead covers extended mind theory, a brief tour through two thousand years of cognitive tool anxiety, and then the crucial distinction between extending a mind and replacing one. That distinction is the axis the rest of this course turns on.

Start with the theory itself, because it's been misrepresented so often that it's worth getting right. In 1998, philosopher Andy Clark and philosopher David Chalmers published a paper asking a deceptively simple question: where does the mind stop and the rest of the world begin? Their answer, developed carefully through a series of thought experiments, was that the boundary is not the skull — it's wherever the relevant cognitive processing is happening. As Andy Clark argues in a 2025 commentary in Nature Communications^[1], humans are and have always been what he calls "hybrid thinking systems" — ones that fluidly incorporate non-biological resources. The mind, on this view, is defined not by biology but by function. If a resource is reliably available, trusted, and doing genuine cognitive work, it is part of the thinking system, full stop.

What extended mind theory does NOT say is worth being equally precise about. It is not saying that your pen is conscious. It is not saying that your smartphone has beliefs. It is not claiming that cognitive tools are the same as biological brains, or that there is no meaningful difference between what happens inside the skull and what happens outside it. The claim is more targeted than that: when you are actively using a tool to think — not just to store or transport information but to genuinely process and extend a cognitive task — that tool is functioning as part of your cognitive system in a way that matters. The tool is not thinking for you; it is thinking with you.

This matters enormously because most of the moral intuitions people bring to AI — the sense that it's cheating, the worry that it "doesn't count," the fear that relying on it makes you less of a thinker — all presuppose the skull-boundary model of mind. They assume there's a pure, unassisted cognitive self in there somewhere that gets contaminated whenever it touches a tool. But as Clark notes in that same Nature Communications piece^[1], this image of ourselves as "nothing but our own biological brains" is a highly restrictive and empirically unsupported self-conception. It is, in his framing, a mistake — one with a remarkably long history.

The history is worth sitting with, because the pattern it reveals is almost embarrassingly consistent. Every major cognitive tool in human history triggered an anxiety that sounds, in retrospect, nearly identical to the anxiety about AI today. The specific technology changes. The fear that it will hollow out some precious, irreplaceable inner capacity does not.

Clark points directly to Plato's Phaedrus, written around 370 BC^[1], where Socrates voices the concern that writing — the new-fangled invention of the age — would have catastrophic effects on human memory. The worry was precise: people would mistake the external record for genuine knowledge. They would think they knew things they merely had written down. They would become cognitively lazy, shored up by a cheap substitute for real understanding, their memories withering from disuse. The fear was that writing would produce a generation of people who appeared knowledgeable but were actually hollow.

Sound familiar? It should. The structure of this fear is identical to today's concerns about AI. And yet — as Clark drily observes — those fears about writing now seem laughable. The invention of writing did not produce cognitively diminished humans. It produced cognitively transformed ones. It enabled the accumulation, transmission, and refinement of knowledge across generations in a way that made entirely new kinds of thinking possible. Mathematics, science, law, philosophy, literature — the things we most value about human intellectual achievement — all depend on writing as a cognitive scaffold. The "pure" oral culture that writing supposedly corrupted was not superior; it was operating with different cognitive tools and could therefore think different thoughts.

Stay with this for one more step, because it pays off substantially. What writing did was not replace memory — it reorganized the cognitive system around a new division of labor. Humans could now offload the verbatim storage of information onto clay, papyrus, and eventually paper, and devote more biological cognitive capacity to interpretation, synthesis, and argument. This is cognitive extension in action: the boundary of the thinking system expanded to include the new technology, and the result was not diminishment but amplification. The fears were not entirely baseless — Socrates was right that people would sometimes mistake external records for genuine understanding. But that failure mode was a problem of implementation, not a reason to reject the tool.

The printing press triggered a parallel wave of anxiety in the fifteenth and sixteenth centuries. When movable type made books cheap and widely available, critics worried about information overload — that the proliferation of texts would overwhelm human discernment, that people would be exposed to dangerous ideas before they had the maturity to evaluate them, that authority and wisdom would be diluted by accessibility. There were genuine concerns about the spread of misinformation through printed pamphlets. Sound familiar again? Every era's cognitive tool anxiety is essentially the same essay with new proper nouns. And in every case, the transformation proved more complex than either the enthusiasts or the doomsayers predicted — new capabilities came alongside new failure modes, and humans adapted, sometimes gracefully, sometimes badly, over time.

Now consider three case studies that bring this closer to home, because the pattern applies not just to ancient or medieval cognitive tools but to technologies within living memory.

The calculator is instructive. When electronic calculators became widely available in the 1970s and affordable enough for classrooms in the 1980s, the debate about whether to allow them in schools was fierce. The concern was that students who used calculators would fail to develop numerical fluency — they would not internalize multiplication tables, would not develop number sense, would come to depend on a device that might not always be available. Some of those concerns were legitimate. Calculators do change the cognitive division of labor between student and tool, and if a student never learns the underlying arithmetic at all, there are genuine gaps. But the overcautious response — banning calculators entirely — confused a question of implementation with a question of principle. Calculators did not make mathematicians worse. They freed mathematicians from the drudgery of lengthy manual computation and allowed them to explore more complex problems. The boundary of what was cognitively feasible in mathematics expanded. The extended mind theory prediction turned out to be correct: offloading well-defined procedural tasks to a reliable tool can liberate higher cognitive functions for more demanding work.

The notebook is a subtler example, and worth dwelling on precisely because nobody has ever worried about it. The practice of keeping written notes — in a journal, a commonplace book, a research notebook — has been treated as unambiguously virtuous for centuries. Francis Bacon kept detailed notes. Darwin's notebooks were integral to the development of his theory of evolution. The scientist or scholar who maintains meticulous notes is considered disciplined and rigorous, not cognitively lazy. Yet functionally, a notebook does exactly what critics worry AI does: it offloads memory to an external system. You trust the notebook to store information you would otherwise have to hold in your head. You use it to extend your cognitive reach beyond what unaided biological memory could manage.

Nobody calls that cheating. And the reason nobody calls it cheating reveals something important: the moral intuition about cognitive tools is not actually about the offloading. It's about whether the offloading is perceived as legitimate given the context and the stakes. The notebook is accepted because it's slow, effortful, and associated with the virtuous act of writing things down. The AI assistant triggers suspicion partly because it's fast, effortless, and associated with the suspicious act of getting someone — or something — else to do the work. But neither of those associations tracks what actually matters, which is what the tool is doing to the thinking and the thinker over time.

The GPS case is the one that gets the most research attention because it has the most visible downside, and it's worth examining carefully. As Clark notes in his Nature Communications commentary^[1], empirical studies have found that GPS apps appear to erode unaided navigational abilities — regular GPS users show measurable differences in hippocampal engagement during navigation compared to people who navigate without assistance. This is the finding that gets cited most often to support the "AI will make us dumber" narrative. And it is a real finding, not a confection. Heavy reliance on GPS does appear to change how people navigate.

But notice what this actually shows. It does not show that the extended cognitive system — human plus GPS — is worse at navigation than the unaided human. It almost certainly isn't; the GPS-assisted navigator gets to more destinations more reliably than the unaided one in almost every scenario. What it shows is that the unaided component of the system has changed — that when you remove the tool, the biological portion performs differently than it did before the tool existed. Whether this is a loss or a reorganization depends on what you think navigation is for. If you believe that the internal capacity for spatial reasoning is intrinsically valuable and worth protecting — for its own sake, or because GPS might not always be available — then the GPS case is a genuine warning about how to use cognitive tools. If you believe that navigation is ultimately about getting from A to B reliably, the GPS case looks rather different.

This is the key distinction the rest of the course depends on, and it's worth stating as precisely as possible: cognitive extension is when a tool expands the reach of a cognitive system, with the biological component remaining engaged and the overall system performing at a higher level. Cognitive replacement is when a tool takes over a function entirely, the biological component stops engaging, and the capacity atrophies in ways that leave the person worse off when the tool is unavailable — or, more subtly, worse off even with the tool, because they've lost the ability to evaluate or guide what the tool is doing.

The GPS case sits somewhere in between, depending on how it's used. Someone who uses GPS to navigate unfamiliar cities while still maintaining their own spatial attention — noticing landmarks, building a mental model of the route — is using GPS as extension. Someone who turns off their spatial attention entirely, stares at the arrow on the screen, and could not find their way back to a place they've visited fifty times without the device — that person is drifting toward replacement. The tool hasn't changed. The usage has.

As the research on generative AI synthesized in a 2025 PMC paper on cognitive extension and atrophy^[2] points out, this is where AI becomes genuinely different from previous cognitive tools in ways that matter. Earlier tools automated well-defined, discrete cognitive procedures: arithmetic computation, information storage, route calculation. These are what the research calls "procedural or retrieval-based" offloading. Generative AI — large language models in particular — is qualitatively different because it can perform integrative reasoning. It can synthesize, argue, evaluate, connect ideas across domains, construct explanations, draft original content. These are not peripheral cognitive tasks. They are central to what we mean by thinking.

This distinction is not a reason to fear AI or refuse it. But it does change the stakes considerably. When a calculator does arithmetic for you, the cognitive territory it occupies is relatively narrow. When a language model drafts your argument, structures your reasoning, and selects your supporting examples, it has moved into cognitive territory that is much closer to the core of what we'd call your thinking. As that same PMC synthesis notes^[2], this marks a fundamental shift in cognitive offloading that presents "unprecedented opportunities for learning but also a challenge that is unique in its nature and scale." The opportunity and the risk are inseparable, which is exactly why the extended mind framework matters more for AI than for any previous tool.

Here's where the philosophical framework pays off practically. If you start from the skull-boundary model of mind — the assumption that real thinking happens inside and AI is always outside — you end up with a binary choice: use AI and compromise your real thinking, or avoid AI and preserve it. That's a bad frame, and it produces bad decisions. You either avoid an extraordinarily powerful cognitive resource out of misplaced purity, or you use it in ways that actually do erode your capacities because you haven't thought carefully about what you're doing.

The extended mind framework gives you a better question. Not "am I using AI?" but "what kind of cognitive relationship am I building with this tool?" Not "is this cheating?" but "is my thinking expanding or contracting?" Not "does this count as real intelligence?" but "is this configuration — me plus AI — making me a more capable thinker over time, or just a more capable output-producer in the moment?"

As the paper exploring AI and human intelligence in PubMed^[3] frames it, the key question is whether AI is extending and thereby enhancing people's minds, or merely enabling them to behave as if they have been cognitively enhanced — to produce intelligent-seeming outputs without the underlying development. That distinction — genuine extension versus mimicry of intelligence — turns out to be the most important thing to understand about AI as a cognitive tool. And it's a distinction you can only make clearly if you start with the right model of what a mind is.

The "it's cheating" intuition deserves one more direct treatment, because it will probably resurface for you even after this argument. The intuition has a real grain of truth at its core: there are contexts where using a tool does undermine the purpose of the activity. If you're learning to multiply and you reach for a calculator before you've internalized the concept, you've short-circuited the learning process. If you're practicing navigation to build spatial skills and you immediately defer to GPS, you've done the same thing. The tool has replaced a process that was supposed to build capacity. In those contexts, the intuition is right — not because tools are inherently bad, but because the goal was the struggle, not the output.

But that's a much more precise claim than "using AI is cheating." It's a claim about specific contexts where the effort itself is the point. And it immediately raises the right follow-on question: which contexts are those? In which domains and at which stages of your development should you deliberately protect the struggle? That question is answerable. The blanket "it's cheating" intuition is not answerable, because it treats all AI use the same regardless of context, domain, purpose, and what the person already knows. The extended mind framework dissolves the blanket prohibition and replaces it with a much more useful, if more demanding, set of discriminations.

Humans are natural-born cyborgs — Clark's phrase from his extended mind work^[1] — and have been since the first person made a mark on a cave wall to remember where the water was. Every generation discovers that this is true, panics briefly, and then integrates the new tool so thoroughly that the integration becomes invisible. The next generation grows up not knowing what the anxiety was about. This is the pattern, and AI will almost certainly follow it.

What's genuinely different this time is the scope and the speed. AI can do integrative reasoning in ways no previous tool could, which raises the stakes for how carefully people engage with it. Understanding the extended mind framework is not just a philosophical nicety — it's the foundation for every practical decision that follows. You can't design a healthy cognitive relationship with AI unless you understand what kind of thing cognition actually is and how tools have always participated in it.

That's the ground this section establishes: your mind has always extended beyond your skull, the tools you use have always shaped what you could think, and the question has never been whether to extend your cognition but how. The next challenge is understanding which cognitive capacities are genuinely at risk when the extension goes wrong — which is precisely where the concept of the hollowed mind comes in.

Sources cited

1. As Andy Clark argues in a 2025 commentary in Nature Communications nature.com ↩
2. As the research on generative AI synthesized in a 2025 PMC paper on cognitive extension and atrophy pmc.ncbi.nlm.nih.gov ↩
3. As the paper exploring AI and human intelligence in PubMed pubmed.ncbi.nlm.nih.gov ↩