In 1998, philosophers Andy Clark and David Chalmers published a short, provocative paper in the journal Analysis that asked a deceptively simple question: where does the mind actually stop? Their answer, built around a thought experiment involving a man named Otto and a notebook, argued that the boundary of a mind isn’t necessarily the skull — that under the right conditions, an external tool can function as a genuine, literal part of someone’s cognitive system, not just an aid to it. Nearly three decades later, the tools people use to store and retrieve their own thoughts have gotten dramatically more capable, talking back, summarizing, inferring things about their users that were never explicitly written down. Clark has already, on his own, applied his original framework directly to this new generation of AI tools. The interesting question isn’t whether the extended mind thesis applies to AI memory — that connection has already been made. It’s whether it actually survives contact with the specific, testable criteria Clark and Chalmers built their argument on.
Scientific Foundation
Clark and Chalmers built their case around a comparison between two people trying to remember the same fact: the address of a museum exhibition. Inga, with an ordinary biological memory, simply recalls that the museum is on 53rd Street and walks there. Otto, who has a memory impairment, carries a notebook everywhere he goes, writing down important information and consulting it when needed; he looks up the museum’s address in his notebook and walks there. Clark and Chalmers argued that, functionally, Otto’s notebook entry plays exactly the role an ordinary belief plays for Inga — it’s reliably available, he trusts it without hesitation, and he acts on it the same way she acts on her memory. Their conclusion, built on what they called the parity principle, was that if an external process does the same job a process happening inside the skull would do, there’s no principled reason to deny it the status of being genuinely, constitutively part of the person’s mind, rather than merely a tool the mind uses. For a resource like Otto’s notebook to count as a real extension of memory rather than just a handy reference, Clark and Chalmers specified it needed to meet four conditions: the resource had to be constantly available and reliably invoked, its contents had to be automatically endorsed without the kind of scrutiny applied to arbitrary outside information, it had to be easily accessible when needed, and — critically — the information stored in it had to have been consciously endorsed by the person at some point in the past, typically because they were the ones who put it there.
Cross-Domain Connection
This isn’t a speculative leap this piece is making on its own. Clark has directly extended his own framework to generative AI, introducing a concept he calls extended cognitive hygiene, a set of practices meant to ensure that cognition delegated to an AI system genuinely contributes to, rather than degrades, a person’s own thinking. Current computer science research is already applying the original four-part test directly to large language model memory architectures, treating Clark and Chalmers’ 1998 criteria as a live, workable framework for evaluating whether a modern AI memory system functions as a genuine cognitive extension or merely as an external database a person happens to consult.
Applied literally, current AI memory tools — the kind that persist facts about a user across conversations, or “second brain” note-taking systems paired with an AI layer — do reasonably well on two of the four criteria. They’re often more constantly available than a physical notebook ever could be, synced across devices, immune to being left at home or lost. And they’re frequently easier to access than flipping through a notebook’s pages, since natural-language and semantic search can surface a relevant memory without the person needing to know exactly where they filed it.
What Remains Undemonstrated
The other two criteria are where the comparison gets genuinely interesting, and where it starts to strain. Automatic endorsement — trusting the stored information without the skepticism you’d apply to a stranger’s claim — holds up reasonably well for content a user directly and explicitly told the system. It gets shakier once an AI memory system starts doing something Otto’s notebook structurally could not: generating its own summaries, inferences, and conclusions about the user, and storing those as if they were equivalent to something the user actually said. That runs directly into the fourth, and arguably most important, criterion: past endorsement. Clark and Chalmers were explicit that Otto’s notebook entries counted as extensions of his mind because Otto himself had written them, the product of his own reasoning and perception at some earlier point. An AI-generated inference about a user’s preferences or personality, produced by the system’s own pattern-matching rather than consciously authored and endorsed by the person it describes, doesn’t clear that bar in the same way — it’s something closer to another party’s opinion about you, filed in a place that happens to be convenient for you to access, rather than a genuine externalized record of your own prior belief. The fact that Clark felt the need to introduce an entirely new concept, cognitive hygiene, specifically when extending his framework to AI, is itself a signal that he doesn’t treat this as a clean, automatic extension of the original 1998 argument — it’s a recognition that something structurally different is happening.
There’s a second disanalogy worth taking seriously rather than glossing over. Otto’s notebook was an entirely passive medium, with no interests, incentives, or interpretive agenda of its own; whatever was written stayed written, exactly as he’d left it. A modern AI memory system is neither passive nor neutral in the same way — it’s operated by a company with its own design choices and incentives, and critically, it doesn’t just store and return information unchanged; it actively synthesizes, reframes, and reinterprets that information every time it’s retrieved. That’s a kind of intermediary standing between “what you once believed” and “what gets handed back to you” that a paper notebook, by its nature, never had to contend with.
Why It Matters
The value of holding the AI memory question to Clark and Chalmers’ original, specific criteria, rather than treating “extended mind” as a loose metaphor, is that it gives a genuinely useful diagnostic tool rather than just a flattering comparison. It suggests a real, practical distinction worth drawing: the parts of an AI memory system built from content a person directly stated and would recognize as their own likely do function as something close to genuine cognitive extension, in the sense Clark and Chalmers meant. The parts built from the system’s own inferences and summaries about the person are doing something categorically different, closer to consulting an outside opinion that’s been given unusually convenient, constant, and trusted access to your daily decisions. Confusing the two, treating an AI’s autonomous inference about you with the same automatic trust you’d extend to your own handwritten note, is precisely the failure mode a careful reading of the 1998 paper would flag as a mistake, not a validation of it.
Human Dimension
There’s something worth sitting with in the fact that a philosophy paper written before most modern AI tools existed already contained, buried in its own fine print, the exact test needed to catch what makes today’s version of the idea more complicated than Otto and his notebook. Otto never had to wonder whether his notebook was quietly editorializing on his behalf. The tools now doing a version of Otto’s job, for millions of people, increasingly do exactly that — and the honest answer to whether that still counts as an extension of someone’s own mind, rather than a new kind of company they’re keeping, is one Clark and Chalmers’ own careful criteria were built to help answer, decades before anyone needed to ask it seriously.
Sources:
1. Notre Dame Philosophical Reviews — “The Extended Mind” — https://ndpr.nd.edu/reviews/the-extended-mind/
2. PhilArchive — “A Challenge to the Extended Mind Hypothesis” (Elaine Jackson) — https://philarchive.org/archive/JACACT-13
3. Medium — “In Defense of Otto and His Extended Mind” (Patrick Buchen) — https://medium.com/@pnbuchen/in-defense-of-otto-and-his-extended-mind-9786db756f2d
4. arXiv — “Doom Researching: A Conceptual Framework for Repetitive AI-Assisted Information Seeking, Cognitive Offloading, and the Illusion of Knowing” — https://arxiv.org/pdf/2607.02723
5. arXiv — “Cognitive Workspace: Active Memory Management for LLMs — An Empirical Study of Functional Infinite Context” — https://arxiv.org/pdf/2508.13171
6. Procedia Computer Science (ScienceDirect) — “The Extended Mind and the Extended Agent” — https://www.sciencedirect.com/science/article/pii/S1877042813037105/pdf
7. arXiv — “The Instrumental Dissolution of Typing: Why AI Challenges the Keyboard Era in Knowledge Work” — https://arxiv.org/pdf/2604.17023
8. PMC (National Institutes of Health) — “Against Strong Ethical Parity: Situated Cognition Theses and Transcranial Brain Stimulation” — https://pmc.ncbi.nlm.nih.gov/articles/PMC5386970/
9. Minds Online — “The What And Where Of Mental States: On what is distinctive about the extended mind thesis” (Karina Vold) — https://mindsonline.philosophyofbrains.com/2016/2016-2/the-what-and-where-of-mental-states-on-what-is-distinctive-about-the-extended-mind-thesis/
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 