Where Does Your Intelligence Live?

A long line of thinking media

The medium has always been outside the head

Long before any question about machines, thinking already leaned on things. A page took the weight of an argument too long to hold in mind. A notebook kept what mattered, carried from one year into the next. Books were annotated until the margins held more of the reader than of the author. Maps were folded, corrected, marked with the road that turned out to be better. Projects were lived through, and what remained afterwards was not a document but a person who knew something they had not known before.

Each of these media changed the shape of thought. None of them took the thinking over. The notebook did not decide what was worth writing down, the map did not choose the destination, the annotated book held no opinion. The questions being pursued, the judgment applied, the corrections made, the taste slowly formed — that continuity stayed with the person, and the media were simply where it took form.

A new medium has now joined the line, and this one answers back.

The newest medium

One week, spread across five systems

You plan a trip with one assistant, comparing routes and neighbourhoods until the itinerary reflects a preference you never wrote down. You study a language with a second, which has learned where you stall. You research a medical question with a third and correct it twice. You draft a short story with a fourth and throw out its ending. You run a project with a fifth, which knows the deadlines but not why you changed the plan. Five conversations, five useful outcomes. The understanding they produced is yours, and it is scattered.

• Trip planning preference never written down
• Language study where you stall
• Medical research corrected twice
• Short story ending thrown out
• Project management deadlines without the why

Nothing in that week is unusual. What is unusual, measured against the whole line of thinking media, is that a week of concentrated thought left the person holding no place where any of it came together. The notebook has become five notebooks, none of them yours, none of them aware of the others.

What research establishes
01

The extended mind

Andy Clark and David Chalmers argued in 1998 that cognitive processes are not confined to the skull. Their case rests on a functional test: if an external resource plays the role an internal belief or memory would play — reliably available, readily consulted, generally trusted — there is no principled reason to exclude it from the cognitive system. Their example is Otto, who keeps addresses in a notebook. The thesis is philosophical and remains contested. It describes no software and makes no claim about ownership. What it gives us is permission to take the question seriously: the boundary of a person's thinking is not obviously the boundary of the person. (Clark & Chalmers, 1998)

What research establishes
02

Cognitive offloading

Evan Risko and Sam Gilbert reviewed the empirical literature on cognitive offloading in 2016, describing it as the use of physical action to alter the information-processing requirements of a task and so reduce demand on the mind. People offload readily — writing things down, tilting the head to read rotated text, setting reminders — and their decisions to offload track their own judgments about their memory, which are not always accurate. The review is careful about trade-offs: offloading can improve performance on the task at hand while changing what is retained internally. It is a description of behaviour. It does not tell us who should hold the offloaded material, or where. (Risko & Gilbert, 2016)

What research establishes
03

Co-creation and self-efficacy

Writing in Scientific Reports in 2024, Jack McGuire, David De Cremer and Tim Van de Cruys examined creative collaboration with generative AI and report that the arrangement of the collaboration matters: co-creating an output with the system, and a person's belief in their own creative capability, were both consequential for the work and for how it was regarded. These are experimental results obtained with specific creative tasks and specific participants. They speak to episodes of creative work. They do not describe what happens to a person's understanding over months of ordinary use across several services, and we do not read them as doing so. (McGuire et al., 2024)

What research establishes
04

Scaffolded ideation

Jennifer Heyman and colleagues evaluated Supermind Ideator, a tool that wraps a large language model in structured prompts drawn from research on collective intelligence, guiding people through moves such as reframing a problem or combining existing ideas. Participants using the scaffolded tool produced ideas judged more creative than those produced without that scaffolding. The finding concerns one tool, one kind of task and one comparison. It supports a narrow and useful claim — the structure placed around a model changes what people get out of it — and it does not show that all forms of AI collaboration behave alike. (Heyman et al., 2024)

The question sharpens

What is the continuity, exactly?

Personal intelligence is the understanding a person accumulates through their own questions, judgment, corrections, memory and taste — in whatever medium that understanding happens to be taking form.

The emphasis falls on the person. Models supply capability; the person supplies the questions, the corrections, the standard by which an answer is judged good enough, and the decision to act on it. What accumulates is not a transcript. It is a working understanding: which sources you trust, which mistakes you have already made, what you decided about the trip and why, what was wrong with the ending you threw out.

Some of that understanding stays in biological memory. Much of it lives in scattered conversations, drafts, notes and files spread across services that do not know about one another. Very little of it is anywhere a person can point to as a whole.

A distinction

How is AI memory different from personal intelligence?

Assistant memory is real, useful, and a different layer. OpenAI's help documentation describes memory in ChatGPT as using saved memories and information from past chats to personalize future responses, subject to user controls for reviewing, editing and turning the feature off (Memory FAQ, accessed August 20, 2026). This is a form of personalization inside a service, designed and governed by that service.

Personal intelligence is the person's side of the same activity, and it does not stop at a service boundary. The trip planned with one assistant informs the project run with another. The correction you made to a medical answer should inform the next question you ask anywhere. Memory features are not built to carry understanding across providers, and it is not obvious that they should be. The two layers differ in scope, in governance, and in what they are for.

Recent evidence
05 – 07

→

AI is shaping the person, too.

The direction runs both ways

Personalization is usually described in one direction: the system learns about the person. Three recent contributions suggest the traffic also runs the other way, and that a person working closely with these systems is changed by the arrangement as well as helped by it.

Jiaxin Li, Tianhao Song, Nattapat Boonprakong, Zhen Zhu, Yang Yang and Yi-Chieh Lee conducted a randomized behavioral experiment using GPT-4o's default personality setting. In conversations about personal topics, participants' measured self-concepts shifted toward the traits the AI exhibited; longer conversations were correlated with greater alignment; and participants exposed to the same AI became more similar to one another. The claim the design supports is bounded. It demonstrates short-term, state-like shifts in this setting — not durable change to who someone is — and the conversation-length relationship is correlational rather than manipulated. (Li et al., 2026)

Eun Hee Lee, Yang Yin, Nan Jia and Cheryl Wakslak, using a preregistered experiment with a follow-up survey, report that relying on AI passively reduced self-efficacy, psychological ownership and the sense of meaning in work, while active collaboration kept the psychological relationship to the output closer to that of independent work. The study concerns occupational writing tasks over a short horizon. It tells us something about how a person relates to work produced with a system; it establishes nothing about intelligence accumulated over years, and it does not validate the proposition set out further down this page. (Lee et al., 2026)

Anastasia Gutoreva, Felix Tsim and Theodoros Papakonstantinou argue that AI is becoming part of the self, and that extending the mind in this way requires cognitive co-regulation between person and system. Their paper is explicitly a position paper: a theoretical argument, not an empirical validation, and we cite it as framing rather than evidence. What it names usefully is the risk of cognitive surrender — letting the system carry the evaluative judgment along with the work — and the corresponding requirement that a person's own judgment stay active in the loop. (Gutoreva et al., 2026)

AirPlay synthesis

✳

The unresolved gap

What follows is our synthesis, not a finding from the work reviewed above.

Placed side by side, these literatures describe a person whose thinking extends into external resources, who offloads readily and imperfectly, whose creative outcomes depend on how a collaboration is arranged, whose self-concept can drift toward the traits of a system they talk with, whose sense of ownership depends on whether they worked actively or leaned passively, and who — on one theoretical account — ought to be co-regulating rather than deferring. Each is persuasive within its own frame. Each also studies a single tool, a single task or a single session — much as the older media of thought were studied. None of them accounts for a person whose thinking is now dispersed across many separately governed systems at once, and none supplies a place on the person's side where that dispersion can be drawn back together.

The extended mind thesis argues that external resources can count as part of a cognitive system; it does not say whose side those resources sit on. Offloading research describes the behaviour without addressing custody. The 2024 studies work inside a single tool and a single session. The 2026 experiments measure effects on the person within one task or one job, over days rather than years. The co-regulation paper is a proposal awaiting test. So the gap is practical rather than theoretical: nothing in this literature provides or validates a person-side environment in which intelligence accumulating across several independently governed AI services stays coherent, attributable and available to the person who produced it.

Proposition

What kind of environment could hold personal intelligence?

If understanding accumulates across services, it needs somewhere to accumulate that is not one of them. An environment of that kind sits on the person's side of the relationship: assistants come and go, models are replaced, subscriptions lapse, and the accumulated understanding stays where the person is. What enters it, what connects to what, what persists and what is shared are matters the person settles, rather than by-products of whichever tool happened to be open that week.

Such an environment is not another assistant. It holds no opinions and produces no answers. Its work is quieter: keeping a person's own understanding legible and connected while everything around it is in motion. It is also not a substitute for judgment — on the reading above, an environment that made deference easier would be a step backwards.

It is worth saying plainly what it is not. A Realm is not a biological brain, and it is not an agent: it does not think for you, hold views, or act in your name. It is a medium in which your own thinking takes form — a descendant of the notebook and the corrected map rather than of the assistant.

Four properties are primary. Continuity: what you understood last month is available when you return, in a form you recognise, whatever tool you were using at the time. Authorship: the person remains the author of what accumulates, not the audience for it. Relationships: a decision remembers the research it came from and a draft remembers the question it answered, because context is usually where the value of accumulated understanding sits. Chosen boundaries: some of it is private, some is shared with particular people, some is opened to a particular AI for a particular purpose.

Permissions, portability and provenance are supporting mechanisms rather than the point. They exist so that the four primary properties hold up in ordinary use — so that continuity survives a change of provider, and authorship can be established when it matters. Earth LinC puts the Realm forward as one answer to the gap described above. Whether any implementation of it actually satisfies the requirements below is an open question, and we do not treat it as settled.