Externalizing knowledge. We no longer know, we know where to ask

Daniel Wegner introduced in 1985 the concept of transactive memory to describe how couples and small groups distribute knowledge instead of duplicating it. You remember the dates, I the names; you keep the accounts, I keep the addresses; together we know what neither knows alone. Memory stops being in a single head and comes to be in the system. With the encyclopedia, then the archive, then Google and now AI, that same distribution has been externalized to technical systems. We no longer know: we know where to ask. The unease any transition of this kind drags with it is structural —what happens when the external thing fails?— and the short answer is that we're not prepared. Which is, in general, normal in societies that discover their dependence the day of the failure.

Transactive memory

Wegner published Transactive memory: A contemporary analysis of the group mind in 1985 (a chapter in Theories of Group Behavior, Springer). The idea operated as an empirical description: stable couples show recall patterns where each member selectively stores what the other doesn't store. The memory of the aggregate is greater than that of any of the parts. It doesn't require telepathy: it requires knowing the division and being able to access the other when needed.

The model spread quickly to work groups. Effective teams aren't the ones that duplicate knowledge but the ones that distribute it. The surgeon knows anatomy, the nurse knows protocols, the anaesthetist knows pharmacology. The group works because each one knows who knows what. Losing a member isn't only the loss of hands: it's the loss of the fragment that member stored.

Wegner and colleagues extended the frame, in 2011, to the human-machine relationship. Sparrow, Liu and Wegner published Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips (Science 333, 2011) with four experiments showing something concrete: when subjects expected the information would be available to them online, they remembered it worse. But they remembered better where it was stored. Memory hadn't disappeared; it had shifted its content. From what to where.

The asymmetry of the change matters: the external system now counts as a member of the transactive memory aggregate. And, like any member, its availability conditions aggregate performance.

The historical chain

It's worth recalling that the externalization of knowledge didn't begin with Google. It's a millennia-old process, and each step displaced something:

• Oral tradition. Knowledge lives in the memory of individuals and is transmitted verbally. Friction is high; depth is maximal; temporal fidelity, low. The Greek bards sang Homer because the Iliad was in heads, not papyri. Whoever knew, knew deeply.
• Writing on a medium. Tablet, papyrus, parchment. Information shifts from the body to the medium. Collective capacity grows massively; temporal fidelity, too. But the media are scarce and expensive. Access is restricted to literate elites.
• The printing press. The friction of production falls several orders of magnitude. Eisenstein, in The Printing Press as an Agent of Change (Cambridge UP, 1979), documented how this step cognitively reorganized Europe between 1450 and 1700. The exact citation, the footnote, the comparison between editions become possible. Intellectual skill changes: the medieval scholastic who mastered thirty authors gives way to the Renaissance humanist who consults three hundred.
• The encyclopedia. Diderot and D'Alembert (1751-1772) systematize accessible knowledge in an alphabetically ordered reference work. General culture stops being memorization; it becomes knowing how to look it up in the encyclopedia. The elites keep memorizing to signal belonging, but the average literate population can operate by consultation.
• The public library. Andrew Carnegie and analogous movements in Europe, the second half of the nineteenth century. The encyclopedic is democratized: the common citizen can consult what was once a privilege.
• The online search engine. Google (1998 onward). Friction falls to seconds. Consultation stops requiring physical travel, leafing through indexes, classificatory literacy. Sparrow, Liu and Wegner documented in 2011 the cognitive effects of the change.
• Generative AI. ChatGPT (2022 onward). You no longer even need to formulate the query well: the system reformulates, asks back, offers a spoken synthesis. The friction approaches a dialogue with another person.

Each step reduces the cost of accessing knowledge, widens the amount accessible, and shifts what counts as personal knowing. In oral tradition, to know was to retain. In print, to know was to handle references. In the encyclopedia, to know how to look up. In the search engine, to know how to formulate a query. In AI, to know how to prompt. Each generation defines knowing by the friction it knows, and judges the previous ones with condescension.

What the average citizen «knows»

What an averagely informed citizen knows operationally without consulting anything in 2026 includes:

• Their own name, current address, date of birth, the names of direct family.
• A small fraction of memorized phone numbers —typically one to five.
• Their concrete job and their usual professional routines.
• General geographic data about their country and region.
• A variable fraction of general culture depending on education —the main historical events, some literary classics, basic geography, operative arithmetic.
• Motor skills —driving, cooking, writing, everyday gestures.

What the same citizen knew forty years ago, in addition to the above:

• Twenty to fifty memorized phone numbers.
• Addresses of close friends and family.
• Encyclopedic data —the capitals of countries, historical dates, basic formulas, unit conversions, a wider vocabulary of the mother tongue.
• Memorized practical information —service timetables, recipes, instructions for common appliances.

The difference isn't that the citizen of 2026 is dumber. It's that their operational memory has gone through a massive redistribution. What once occupied mental space now occupies space in external systems. There are gains —a massively greater aggregate capacity— and losses —fragility in the face of external systems going down.

Cognitive offloading

Evan Risko and Sam Gilbert systematically reviewed the literature in Cognitive Offloading (Trends in Cognitive Sciences 20(9), 2016). Their frame distinguishes two types:

1. Material external offloading: notes on paper, reminders in the calendar, lists. An old, well-studied solution. 2. Digital external offloading: apps, search engines, generative models. A recent solution, under active study.

Their central conclusion, comparing the two types: the cognitive effects are qualitatively similar —reduction of memorization effort, redistribution of the cognitive resource— but the scale is very different. Material offloading had significant friction —going to the notebook, leafing, finding the note. The digital kind has almost no friction. The amount of offloadable information multiplies, and so does the frequency with which it's done.

Yifan Hu and colleagues, in Cognitive Offloading in Short-Term Memory Tasks: Trust Toward Tools as a Moderator (International Journal of Human-Computer Interaction, 2025), documented an important moderator: trust in the tool. When subjects trust the reliability of the external tool more, they offload more —even when this harms later recall. The operational consequence: the perceived quality of the AI provider modulates how much offloading users do. Models perceived as highly reliable generate more dependence, not less.

The asymmetry of failure

Here comes the unsettling point. Externalization is efficient in the normal regime. In the exceptional regime, it leaves the user in a qualitatively different position.

Let's compare two people who need, in an emergency, to remember a relative's phone number:

• Person A, of a generation that memorized numbers. They make an effort, it's hard, they didn't have it remembered, but maybe they retrieve it through associations. The capacity exists, even if rusty.
• Person B, of a generation that never memorized. They have no latent capacity. The mental search is structurally empty. Without the device, there's no internal route.

The difference isn't one of degree, it's one of kind. Person A has a degraded but functional system. Person B has an absent system. The efficiency of externalization in the normal regime is paid for, in the exceptional regime, with a total lack of backup.

The same applies at higher levels. The professional who externalizes the drafting of reports keeps the capacity for judgement over the reports. The professional who externalizes the drafting and the judgement is left dependent for both. If the system goes down, the first produces worse reports; the second produces no reports.

The conversation that isn't being had

This is personal opinion but documentable. The social discussion about the externalization of knowledge has polarized into two sterile positions:

1. Retrograde nostalgia. «People used to know things». True in parts, exaggerated in others, useless as a policy guide. Those who profess it usually memorized what they had to memorize and disdain what younger people have learned to handle. 2. Lazy techno-optimism. «What matters isn't knowing, it's knowing how to access». True in parts, also exaggerated, equally useless. Those who profess it usually haven't thought about what happens when access fails, nor about what things require memory stock to be thought at all.

The serious conversation would be in the middle and would be had with three questions:

• Which parts of knowledge deserve to keep being memorized and why? The answer varies by subject, profession and country, but the question is legitimate.
• What material conditions does externalization need to be sustainable? Provider redundancy, guaranteed access, maintenance of a minimal backup skill.
• What kinds of thinking are only possible with significant memory stock? There's literature on this in pedagogy and cognitive psychology. Improvised synthesis, the unexpected connection between domains, the recognition of distant patterns —all require material accessible in the head, not on the screen.

These three questions aren't being asked at the scale they'd require. They're asked in university corridors, in some specialized blogs, in academic books that circulate little. The mainstream public discussion keeps swinging between nostalgia and enthusiasm.

A possible middle position

I have no closed answer for the three questions. I do have an operational intuition: externalization is an instrument, not a virtue or a vice. What matters is what gets externalized, what gets kept, and what conditions are met for the combination to be sustainable.

Applied to generative AI: there are tasks where externalizing makes sense and the loss is marginal —writing routine emails, looking up specific data, operational translations. There are tasks where externalizing has a significant cognitive cost and the accumulated loss is important —original synthesis, professional judgement, sustained reflection on complex topics. The reasonable decision is selective, not global.

The aggregate problem is that the selective decision requires reflection most users don't do. By default, everything gets externalized, because the friction is minimal and the immediate benefit is clear. The aggregate consequence at ten years isn't evaluated at the moment of each decision.

What an educator might think

Educating in this landscape would require things that the schools and universities of 2026 rarely do:

• Teaching how to evaluate when it's worth memorizing and when it isn't.
• Keeping mnemonic training for critical fragments —the fundamentals of the field, operative vocabulary, pivot dates and concepts.
• Practising regularly without assistance to keep a backup capacity.
• Explicitly discussing what's delegated and why.

This isn't a retrograde proposal. It's a proposal of education for the real regime, which includes massive use of externalization but also includes —and should include— maintenance of one's own capacity. The educational systems advancing in this direction do so experimentally and in isolation. Most keep swinging between banning the tools and allowing them with no framework.

Definitions

• Transactive memory (Wegner, 1985): a system of distributing knowledge among members of a group or between a subject and external tools; each part stores selectively and the aggregate works because each one knows where what they need is.
• Cognitive offloading (Risko & Gilbert, 2016): the practice of delegating parts of cognitive work to external supports —notes, devices, systems; it reduces effort and, simultaneously, reduces later recall.
• Google effect (Sparrow, Liu & Wegner, 2011): the documented cognitive effect by which subjects remember information worse when they expect it will be available online, but remember better where it's stored.
• Memory stock: the amount and quality of knowledge internally accessible in the subject's head, retrievable without external consultation; complementary to the stream of information accessible through external systems.
• Asymmetry of failure: the qualitative difference between the subject's performance with external systems available (high) and without them (not degraded but structurally empty in what was fully delegated).

References

• Eisenstein, E. L. The Printing Press as an Agent of Change. Cambridge University Press, 1979.
• Wegner, D. M. «Transactive memory: A contemporary analysis of the group mind». In Theories of Group Behavior, Springer, 1985.
• Carr, N. The Shallows: What the Internet Is Doing to Our Brains. W. W. Norton, 2010.
• Sparrow, B.; Liu, J.; Wegner, D. M. «Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips». Science 333(6043), 2011.
• Risko, E. F.; Gilbert, S. J. «Cognitive Offloading». Trends in Cognitive Sciences 20(9), 2016.
• Hu, Y. et al. «Cognitive Offloading in Short-Term Memory Tasks: Trust Toward Tools as a Moderator». International Journal of Human-Computer Interaction, 2025.

Going deeper

• Hutchins, E. Cognition in the Wild. MIT Press, 1995. The foundational frame of distributed cognition, extending Wegner's idea to complete socio-technical systems —aircraft cockpits, navigation teams, operating rooms.
• Clark, A.; Chalmers, D. «The Extended Mind». Analysis 58(1), 1998. The strong philosophical thesis: external systems can be a literal part of the mind, not just support it; a frame for rethinking what counts as one's own memory.

You might also like

• Knowledge as flow. From stock to stream
• Progressive dependence. Dependence isn't decided, it's discovered
• Learning without remembering
• The degradation of knowledge

Elsewhere

• Supporting Cognition With Modern Technology: Distributed Cognition Today and in an AI-Enhanced Future (PMC)
• A Review of the Negative Effects of Digital Technology on Cognition (arXiv)