Listen: https://serve.buildtall.systems/28565055e9e8f19279c965fd49f1b40749b74c73a1431546bf21432e27a29619.mp3
Source: https://serve.buildtall.systems/e172efe1eee27418aa3bf847b2d8498335c9f4727f723e64219e97af0b161c22.pdf
This episode covers Large-Language Models as a Cognitive Virus, a paper that treats large language models as a possible self-propagating pattern in human thinking and asks how their spread could reorganize the link between language, cognition and technology.
The paper's central question is whether large language models reorganize the coupled system of language, cognition and technology, rather than simply changing how people write. It builds an analogy between these models and viruses, drawing on computer viruses, cultural evolution theory and the idea of cognitive offloading, and then sets out a population model of how people move between three states: uncoupled or weakly coupled use, autonomous coupled use, and persistent dependence on the tool. This is a theoretical modelling paper. The model is a compartment model borrowed from epidemiology, and the text describes no new experiment or dataset. Its numbers come from illustrative parameter choices, and its evidence on education, writing and programming comes from outside studies that the paper summarizes without giving their full designs. Its strengths are its candour about its simplifications and a clear analysis of tipping, runaway change and hysteresis, along with a useful separation between AI as scaffolding and AI as substitution. Its limits are substantial. The model is mean-field, meaning it averages over the population instead of modelling individual contacts, the competence scores are stated to be an illustrative assumption, and the model does not set a time scale for change. The text provided does not include a funding statement or a data availability statement. According to OpenAlex, no work in the index matches this document's title and authors as of October 12, 2026, so OpenAlex has no record of the document, and this episode makes no claim about how often it is cited.
The paper's introduction begins with language, which it treats as a major evolutionary transition that created a new system of inheritance and collective memory. Languages, the authors argue, can be viewed as population-level cultural systems with partially autonomous dynamics. They diversify, compete, hybridize, spread, and sometimes go extinct, in patterns the authors compare to species in ecological communities. Their persistence depends on social learning, the process by which people acquire words, grammar and conventions by observing and interacting with others, especially in childhood. Because linguistic structures propagate by entering developing minds and recruiting their speakers as new carriers, the authors note that language has sometimes been described as a viral entity.
The paper then traces the idea of the virus-like information structure through two strands. The first is computer viruses, self-propagating programs that enter a host system, modify its operation and use it to spread further. The second is cultural evolution theory, which holds that socially transmitted information can vary, persist differentially and be inherited, whether framed through gene-culture coevolution or through the concept of memes. Together, the authors say, these traditions point to a broader class of entities, which they call "viruses of the mind", whose evolution depends not on a particular material substrate but on their capacity to reproduce and persist across biological, cognitive or computational hosts.
In language, the paper says, what spreads virally are patterns, from words to ideas, that reproduce through imitation, mutate through use and persist through brains, media, institutions and technologies. This leads to the idea of the extended mind, in which cognition is distributed across minds and external structures that do not merely store information but reshape cognitive tasks. The authors describe such artifacts as either complementary, strengthening capacities beyond their immediate use, or competitive, improving performance while potentially weakening the underlying skill. Language itself is presented as a two-sided cognitive technology: it externalizes, stabilizes and recombines thought, while linguistic categories and narratives shape perception and reasoning. Because language is adapted over time to the demands of its users, the authors conclude that its propagation both extends cognition and transforms the minds that depend on it.
The introduction argues that communication technologies work in the same reciprocal way. It says the Internet and instant messaging blurred the distinction between speech and writing, and that mobile phones encouraged abbreviated, compressed expression, which the authors read as reflecting linguistic flexibility rather than declining literacy. Social media, in their account, reorganized the structure and social selection of communication by combining text, hyperlinks, multimodal expression and algorithmically selected content. The paper's point is that these technologies do not merely carry language; their affordances reshape how it is produced, circulated, interpreted and selected.
Large language models, the authors write, represent a further transition. Unlike earlier media, they "actively participate in" human expression by producing, reformulating and evaluating it, rather than only constraining its form or circulation. The paper therefore frames the central question not as whether LLMs alter language use, but whether they reorganize the coupled system formed by language, cognition and technology. The comparison, the authors say, may help identify which cognitive capacities these systems amplify, which they displace, and under what conditions sustained reliance may produce cognitive reorganization or erosion. The paper places LLMs within a feature shared by earlier information technologies to varying degrees, which it calls cognitive offloading: the delegation of cognitive operations to external tools and representations.
To bridge natural language and the LLM era, the introduction names two technological precursors, computer viruses and programming languages. The authors recall that computer viruses reproduced key features of biological viruses, being compact informational structures that exploit host machinery and evolve to evade detection, and that they triggered an arms race with antivirus software and shaped cybersecurity and computer immunology. Programming languages, by contrast, spread across computers, institutions and communities, became an operational substrate of the digital revolution and, alongside neural networks, paved the way for artificial intelligence. Both depend on technological hosts and reshape their environments. The paper's summary of the match is that LLMs combine aspects of each: like programming languages they provide a new human-machine interface, and like computer viruses they propagate through users and digital ecosystems. The authors add that their capacity for cognitive offloading is unprecedented.
The paper states its method. It treats LLMs as engines of cognitive and social change whose rapid diffusion is transforming human capacities. It first compares LLMs with infectious agents, then develops an explicit population model of their propagation, drawing on epidemiological mathematics that has also been applied to technology adoption. The model describes transitions among host-coupling states, which the authors link to an illustrative measure of cognitive competence. The aim is to distinguish the consequences of different coupling regimes, to reveal the underlying bifurcation structure, and to explore possible interventions. The authors are explicit that the model does not identify any single entity as the viral analogue. LLM ecosystems contain culturally transmitted practices and content alongside technologically evolving model lineages, and these need not coincide. The analogy concerns instead a broader feedback loop in which persistent technological lineages are instantiated in external machinery, modify their human host environment and thereby influence their own propagation. The paper notes that this loop is already partly present in conventional and adaptive software, but that LLMs deepen it by becoming integrated into cognitive processing itself.
Finally, the introduction stresses that the viral analogy does not imply that LLM-human interactions are intrinsically parasitic. The authors point out that biological viruses range from pathogens to mutualists and evolutionary partners, and that the effects of LLMs depend on how they are used. According to the paper, LLMs can enhance exploration, access to expertise and productivity, but can also promote offloading, dependence and loss of competence, while individual benefits may not scale to the population level. The analogy is useful, the authors argue, because LLM use can generate persistence and transmission loops, both through culturally propagating patterns of use and through model lineages sustained and modified by technological and economic feedbacks. The model that follows, the introduction says, focuses on one part of this ecology: transitions among human states of cognitive coupling.
The paper now sets out its population model, which tracks the share of people in each of three states. The first is uncoupled or weakly coupled users, who make little or no use of LLMs and whose thinking is spread across the brain and a varied information environment of books, writing, computers, Wikipedia and the internet. The second is autonomous coupled users, who use LLMs but keep reading, writing, reasoning and checking claims, and who keep drawing on other sources. The third is persistently dependent users, for whom the model has become the main way they carry out thinking tasks, replacing other supports rather than adding to them.
Movement between these states is governed by a small set of rates. Exposure through social, institutional or platform-mediated LLM practices moves people from uncoupled to autonomous use at a rate the paper calls lambda. People who abandon the technology return to the uncoupled state at a rate rho. Regular use turns into dependency at a rate mu, and training, verification habits or deliberate cognitive friction bring people back to autonomous use at a rate sigma. The authors say these rates "coarse-grain heterogeneous individual processes into population-level transitions rather than representing single psychological mechanisms." They are summaries of many different individual processes, not measurements of single mechanisms.
The model borrows its mathematics from epidemiology, where classical models describe the spread of biological agents through transitions between host states. Similar models have been used to describe technology adoption, where people adopt a tool because others around them already use it, and multi-compartment models have been used for drug and social media addictions. The authors are explicit about what the model leaves out. It does not represent how the model or program lineage reproduces or evolves. Instead, it captures one epidemiological layer of the wider LLM ecology, the transitions among human hosts and their coupling with the technology. In their words, "the fact that the state variables describe hosts does not imply that a host state is the pathogen." They add that
the equations model the epidemiology of coupling and therefore do not, by themselves, determine the identity of the viral analogue.
The model includes a nonlinear term that the authors describe as cooperative, or Allee-like. An Allee effect is a situation where a behaviour becomes easier to sustain as more people practise it. The assumption here is that independent, LLM-free cognitive practice is socially reinforced, and that cultural expectations rewarding independent reasoning work better when many people are autonomous. The term depends on the square of the uncoupled share, which captures this positive frequency dependence. It also depends on the share of regular users, so the restoring effect applies only to people who are already engaged with LLMs. The paper treats the transmission term as an effective measure of social or institutional pressure on hosts to adopt the technology.
The authors acknowledge that their approach is a simplification. Mean-field approximations, which assume that individuals sample the population's overall frequencies rather than specific contacts, are described as providing "a valuable, analytically tractable baseline for identifying the mechanisms underlying collective behavior." Real social and technological networks, though, are heterogeneous and correlated, and the way people are connected can shift epidemic thresholds and the outcomes of immunisation. The paper also points out that communication structure could shape collective behaviour in networks of LLM agents, particularly when people's cognitive states reshape their social ties. The authors present the current model as a reference point for later network and spatial extensions.
The cooperative component has precedents in social and ecological models, and the paper relies on threshold models of collective behaviour. These show that frequency dependence can produce sudden, critical transitions in human groups. The paper notes that social conventions can undergo genuine tipping, with a minority of the group being enough to move the population from one convention to another. On this view, a shift in LLM use could be driven by a committed minority rather than a majority.
A key structural choice is that dependency arises mainly through regular use, not directly from being uncoupled. The authors argue that these simplifications isolate the interaction between contagion-like technological adoption and the collective protection of cognitive autonomy. This, they say, allows them "to identify conditions for abrupt and potentially irreversible population-level transitions."
The paper also places this model within a broader view of human cognition. It notes that offloading mental work onto external tools is normal and often useful. Reading and writing are the paper's main example: literacy reorganises existing neural circuits, and written symbols extend memory and make new kinds of reasoning possible. External support can either build new internal competence or stand in for it. The authors cite recent work on LLM-assisted writing, which found that "increased externally supported performance can be accompanied by reduced cognitive engagement, recall, or sense of authorship."
This leads to the distinction between scaffolding and substitution. Scaffolding lowers immediate cognitive demands while keeping or raising a person's later ability to do the task. Substitution removes the need to perform the underlying mental operation at all. The paper links these to complementary and competitive cognitive artifacts from earlier work. An LLM can act as a tutor whose arguments the user rebuilds, challenges and checks, or as a substitute that does the synthesis, evaluation and writing for the user. The paper concludes that
the relevant distinction is therefore not simply between using and avoiding LLMs, but between forms of coupling and, ultimately, what cognitive capacities remain when the tool is withdrawn.
In the model, the move from regular use to dependency is the minimal representation of this shift from complementary use toward persistent substitution.
The paper then turns to the long-term regimes the model can settle into. The three shares must add up to one, so the dependent share follows from the other two. The authors illustrate the model's bifurcation structure with the equilibrium share of uncoupled users plotted against the transmission rate. For a range of transmission values, the model is bistable: a fully uncoupled state and a coupled state are both stable, separated by an unstable middle branch. As transmission rises past one threshold, the system jumps abruptly into the coupled regime. Recovery does not happen when transmission falls back to that same level. It only happens after transmission drops below a lower threshold, which produces hysteresis and what the authors call technological lock-in. Past the tipping point, the uncoupled share falls sharply while regular use and persistent dependency rise rapidly, which the authors describe as population-level cognitive offloading. The illustration uses a return rate of 0.10, a cooperative strength of 0.40, a dependency rate of 0.20 and a recovery rate of 0.10. Using the dependent share as the remainder, the dynamics of the model are examined next.
The paper then asks what this shift could mean for the competence of the people involved, and its answer is conditional. Cognitive competence is defined narrowly as
the capacity available to the human when external support is removed, rather than the total capability of the coupled human-AI system.
The question is how much a person can still do unaided, not how well they perform with the tool in hand.
To make this concrete, the authors give each group a relative competence score. Fully autonomous people score 1, people in regular use of the tool score 0.5, and dependent people score 0.1, so competence falls as dependence grows. The authors are explicit that this is a choice of illustration:
This ordering is an illustrative modelling assumption, not a general claim about LLM use.
They add that in scaffolded or augmentative uses, coupling could leave competence unchanged or even increase it. The drop described here belongs to one assumed pattern of use, not to the technology in general.
Averaging these scores across the population gives a single figure for the community's competence. That average is an increasing function of the share of people who are still fully autonomous, so it inherits the same tipping behaviour as the adoption dynamics. The paper calls this a hysteresis: a system that changes gradually can jump, and it does not jump back at the same point. For the parameter set used for the main illustration, the mean competence of the coupled population is about 0.233. When the pressure to adopt the tool passes the upper threshold, average competence falls from 1 to roughly 0.425.
The reverse path is different. If the pressure is lowered again, the coupled state holds on until the lower threshold, where average competence is about 0.617. Only then does it recover to unity. The size of the drop depends on that coupled mean and on the ratio of two model parameters at the tipping point. The authors also note that the two rates governing the model do not move the tipping points themselves, but they set the coupled mean and so the cost of occupying the dependent state. In the authors' words,
The bifurcation therefore converts a smooth change in the pressure to adopt LLM-mediated cognition into a discontinuous and history-dependent change in population-level cognitive competence.
The authors also flag a limit.
The minimal model does not, however, determine how rapidly the transition unfolds in real time.
It shows that a jump is possible under slow changes in conditions, but it says nothing about how quickly a real population would move.
To visualise this, the authors describe the competence average as moving over an effective landscape, a potential with two valleys. The authors put it this way:
The dynamics can be pictured as a marble moving over the landscape until it settles in a valley.
At low pressure there is one valley, at high competence, representing autonomy. Past the lower threshold a second valley appears at lower competence, representing offloading. Between the two thresholds either valley can hold the population, depending on its history. As the pressure rises, the offloading valley deepens. At a transmission level of about 0.420 the two valleys have equal depth, and beyond that point the offloading state becomes increasingly favoured.
The passage opens by returning to the landscape picture. The paper describes a marble resting in a valley of high autonomy. When the pressure from LLM use rises past a critical point, that valley vanishes and the marble rolls abruptly toward a low-autonomy minimum. This is a runaway transition: greater reliance on LLMs promotes further cognitive offloading, which in turn reinforces that reliance. Recovery is harder. The marble must cross a barrier, or the pressure must fall far enough that the offloading valley disappears at a second, lower point. The paper summarizes this as an intuitive picture of tipping, runaway change, and hysteresis, meaning that the path a system has taken determines where it can go back to.
The paper then turns to what it calls cognitive immunization. It is "understood here not as preventing contact with LLMs, but as preserving resistance to harmful substitutive or dependency-producing coupling while allowing beneficial forms of human-AI integration." In other words, the goal is not to keep people away from the tools. The goal is to keep the tools from replacing reasoning that people would otherwise do themselves.
The first result the paper highlights is an asymmetry between prevention and reversal. If the population starts close to the fully uncoupled state, invasion is prevented as long as the transmission pressure stays below an invasion threshold. That threshold is set by two rates: the rate at which regular users return to uncoupled cognition and the strength of the collective reinforcement of autonomous thinking. Once the system has moved into the coupled state, however, pushing the pressure below that invasion threshold is not enough, if the system sits in the bistable region, where two stable states can coexist. The coupled state persists until the pressure falls below a lower threshold that depends on the same two rates. Reversal therefore requires a larger reduction in pressure than prevention did. The paper states that
This prevention-reversal asymmetry is the direct dynamical consequence of hysteresis and provides a simple mechanism for technological lock-in.
The return rate has what the paper calls a particularly clear protective role. Raising it lifts the invasion threshold and moves the system toward a boundary at which the lower threshold disappears. Past that boundary, "bistability is lost and the transition becomes continuous." The paper's examples of interventions that raise this return rate include protected unaided tasks, deliberate periods of disengagement, maintenance of non-LLM skills, and attractive non-LLM alternatives. Such measures, the paper argues, can "modify not only the position of the tipping point but the qualitative form of the transition itself."
The effect of collective reinforcement is more subtle. Strengthening it makes invasion harder when most of the population is uncoupled. But beyond the point where it exceeds the return rate, it also widens the gap between the two thresholds, which enlarges the region where the system can be trapped in either state. The paper warns that "Strong collective protection can therefore stabilize autonomy" while also increasing path dependence once the autonomous fraction has been substantially depleted. The implication is that the aim is not simply to maximize collective reinforcement, but to strengthen it together with sufficiently strong return processes, so that protection does not come at the cost of a broad hysteretic region.
Two further parameters play a different role. They do not move either threshold. They control the makeup of the coupled population, setting the share of users who end up in a dependent state. At equilibrium that share is determined by the balance between the two parameters, so lowering one or raising the other reduces the number of dependent users without necessarily reducing overall LLM adoption. The paper proposes metacognitive training, verification requirements, periodic unaided practice, task designs that preserve active reasoning, and mechanisms that help people recover from dependency as ways to act on these parameters.
The model thus separates two levels of intervention. Changes to the pressure, the return rate, and the collective reinforcement reshape the population-level landscape and decide whether bistability and hysteresis are possible at all. Changes to the other two parameters mainly determine the cognitive burden of the coupled state. As the paper puts it,
In this sense, immunization can act either by preventing a collective transition or by reducing the probability that ordinary LLM use develops into persistent dependency.
Immunization in this sense "need not imply low overall LLM use." It can instead mean shaping the coupling so that high adoption stays compatible with verification, active reasoning, autonomous alternatives, and recovery from substitutive dependence.
The discussion opens by returning to 1960, when Joseph Licklider anticipated a future in which the relation between humans and computers would move beyond simple tool use. He wrote:
Man-computer symbiosis is probably not the ultimate paradigm for complex technological systems. It seems entirely possible that, in due course, electronic or chemical machines will outdo the human brain in most of the functions we now consider exclusively within its province.
The authors say that more than six decades later, LLMs made that quote especially concrete, and that their importance lies less in what they can do than in the new forms of coupling they create between human and artificial cognition.
At one extreme, the paper suggests, AI could become an “exocortex”, an external extension of the mind. Such a system could search large bodies of knowledge while leaving interpretation and high-level judgment to the human researcher. The authors argue that this could greatly accelerate the accumulation and transmission of cultural and scientific knowledge.
The authors place this in a longer history. Humans have always externalized parts of cognition into language, writing, institutions, and technology, and they describe that externalization as the engine of rapid, culturally driven change. AI, in their account, is a new step in this process, one in which external cognitive machinery becomes increasingly active and agent-like. As earlier studies have suggested, this might be a new class of synthetic evolutionary transition.
The model, the authors argue, shows that the consequences of this transition cannot be understood from individual LLM use alone. Two social processes interact. The first is adoption. New ways of using LLMs spread because people learn from other people, because workplaces and schools adopt them, and because successful practices are copied, so adoption has a contagion-like component. The second is that autonomous cognition is itself socially sustained. Education and institutions do more than transmit information. They create environments in which reading, writing, reasoning, verification, and independent problem solving are repeatedly practiced and rewarded, and when these practices are common, they reinforce each other.
The interaction between these two processes creates the central feedback in the model. As cognitive delegation becomes more widespread, the social environment that supports autonomous reasoning can weaken, and as that environment weakens, delegation becomes still easier and more attractive. A gradual rise in adoption can therefore produce a disproportionate collective response. Beyond a critical point, the loss of autonomy becomes self-reinforcing, and the population can move rapidly toward much stronger cognitive offloading and lower cognitive competence. The authors are careful on this point. In their words,
The important result is not that such a runaway must occur, but that it can occur under plausible forms of social learning and cooperative reinforcement.
Once such a state is established, returning conditions to where they were before may not be enough to restore the previous state. As the paper puts it,
Prevention can therefore be considerably easier than reversal.
This possibility should not obscure the substantial benefits of LLMs. The distinction, the authors stress, is not between using and not using AI, but between forms of coupling that extend human competence and those that replace the cognitive operations through which competence is maintained. The evidence already points to both possibilities, and the direction of the effects depends strongly on how the human-AI coupling is built. In a randomized study, a guided “think first, ChatGPT later” protocol produced higher later independent creativity than unrestricted use of ChatGPT.
The paper also reviews mixed results from other settings. Generative AI can increase productivity and provide powerful cognitive assistance, but knowledge workers also report reduced effort to think critically when their confidence in AI is high. In education, unrestricted access to generative AI can improve performance while the tool is available, yet reduce unaided performance afterward, while pedagogically constrained AI can substantially mitigate this effect. Related results indicate poorer comprehension or retention when students rely on LLMs without complementary activities such as note-taking. In programming, novice users can struggle to understand and critically evaluate AI-generated code, which creates conditions for automation bias and superficial competence.
These findings, the authors say, reinforce the distinction between AI as scaffolding and AI as substitution, and they change what “cognitive immunization” should mean. At the population level, immunization does not mean preventing contact with AI. It means preserving the practices and institutions that keep human cognition active: unaided problem solving, verification, critical discussion, periods of deliberate disengagement, maintenance of non-AI skills, and educational designs in which the model supports rather than completes the cognitive task.
Once persistent dependency has developed, the paper argues, educational design alone can no longer address the problem. Emerging work on problematic LLM use points to mechanisms related to loss of control, emotional regulation, cognitive biases, and habitual reliance. That suggests that behavioral self-regulation and, in more severe cases, psychological interventions such as cognitive-behavioral therapy may become relevant. In this way the immunization analogy runs from prevention at the cultural and institutional level to recovery at the individual level.
The authors list several extensions that would bring the minimal model closer to real human-LLM interactions. These include replacing the population-level description with heterogeneous interacting agents, letting cognitive autonomy vary continuously rather than splitting users into sharply separated classes, and adding feedbacks through which cognitive change alters later adoption and dependence. Such a formulation would also soften the boundaries between humans, LLMs, and hybrid human-machine systems, which may occupy a broad continuum rather than discrete categories. If agents are persistent patterns of information and activity distributed across biological, artificial, or hybrid substrates, then what spreads may be patterns of cognitive organization, including perspectives that shape how agents represent the world and themselves. On this view LLMs would be more than passive tools for offloading, since they could progressively reshape the cognitive structures through which users interpret and act on the world. In the paper's framing, the “cognitive virus” is a persistent technological lineage embedded in a larger ecology of cultural replicators, human host states, institutions, and technical infrastructure.
The authors acknowledge a further limitation: human behavior is not yet explicitly coupled to the spreading process. Such feedbacks can reshape epidemic thresholds and can even generate higher-order critical phenomena. They point to Diverse Intelligence, an emerging field that treats intelligence and agency as graded capacities of systems operating across multiple spatial, temporal, and organizational scales, and not only as properties of brains or particular biological substrates. They also point to multilayer-network theory as a formal way of representing the interacting behavioral, technological, and social processes through which such feedbacks propagate. Together, they suggest, these approaches offer a systematic route beyond the present mean-field description, which averages over the population. Future models could treat autonomy, dependence, and agency as continuous and coevolving properties of distributed human-machine patterns, and ask under what conditions transient interactions become self-maintaining forms of cognitive organization.
The paper's main takeaway is that the effects of LLMs depend on how the human-AI coupling is built, and that a gradual rise in adoption can, under plausible social learning and collective reinforcement, produce a sudden and hard-to-reverse shift toward offloading and lower unaided competence, so that preventing such a shift is easier than reversing it. The paper is open about what it leaves unresolved. The model is minimal and population-averaged, it does not say how quickly a real population would move between states, and its competence ordering is an illustrative assumption rather than a general claim about LLM use. It also does not identify which entity serves as the viral analogue, and it notes that human behaviour is not yet explicitly coupled to the spreading process, which could change epidemic thresholds. The authors point to heterogeneous interacting agents, continuously varying autonomy, network structure and diverse intelligence as directions for future work that could test whether these transient interactions become self-maintaining forms of cognitive organization.

