A Technical Review of a Provocative New Model of Human–AI Interaction

A provocative new research paper asks whether large-language models should be understood not merely as tools, but as entities that spread through society, attach themselves to human cognitive processes, and potentially create persistent dependence.

The paper is titled “Large-Language Models as a Cognitive Virus”. It was submitted to arXiv on September 3, 2026, by Ricard Solé, Giulio Ruffini, Francesca Castaldo, Marco Tuccio, Luis F. Seoane, Manlio de Domenico, Santiago F. Elena, David C. Krakauer, and Michael Levin.

The authors are affiliated with the Complex Systems Lab at Universitat Pompeu Fabra; the Catalan Institution for Research and Advanced Studies; the Institute of Evolutionary Biology in Barcelona; the Santa Fe Institute; the Barcelona Computational Foundation; Neuroelectrics; the Okinawa Institute of Science and Technology; the University of Padua; the Italian National Institute for Nuclear Physics; the University of Valencia; Tufts University; and Harvard University’s Wyss Institute for Biologically Inspired Engineering.

Although the title is intentionally alarming, the paper does not claim that large-language models are literally biological viruses or that all use of artificial intelligence is inherently harmful. Instead, it uses concepts from epidemiology, evolutionary biology, cultural evolution, and complex-systems science to examine how AI use spreads, how it becomes embedded in human behavior, and under what conditions it might weaken—or strengthen—human cognitive capability.

The paper’s fundamental insight is that the consequences of AI depend less on whether people use it than on how they become cognitively coupled to it.

An AI system can function as a substitute that gradually displaces human competence. It can also function as a scaffold that helps people develop greater competence. At its best, it can become a cognitive collaborator that amplifies human capabilities while leaving human purpose, judgment, creativity, verification, and responsibility intact.

That distinction is essential as society decides how artificial intelligence should be introduced into education, business, government, and everyday life.

Why Compare an LLM to a Virus?

A biological virus cannot reproduce independently. It must enter a host and use the host’s cellular machinery to replicate. A computer virus similarly requires a computing system through which it can execute, reproduce, and spread.

A large-language model also depends upon an extensive host environment. It requires data centers, electricity, computing hardware, software platforms, telecommunications networks, organizations, developers, and human users. Its outputs spread through documents, software, education, business processes, social media, and interpersonal communication.

As people observe others using AI successfully, they are more likely to adopt it themselves. Schools and businesses institutionalize its use. AI-generated practices and content circulate through society. Some of that content becomes part of the data environment used to develop later AI systems. Human beings, institutions, and technological infrastructure therefore form a feedback system through which AI use propagates.

This is the limited but meaningful sense in which the authors apply a viral analogy.

The analogy should not be carried too far. An LLM does not ordinarily possess an independent biological drive to survive or reproduce. It does not autonomously build data centers, purchase processors, connect itself to organizations, or compel people to use it. Human beings and institutions make those decisions.

The paper acknowledges this limitation. Its mathematical model does not simulate the reproduction or evolution of an LLM itself. It models changes in the human population as people move among different states of AI use and cognitive dependence.

The virus is therefore not a clearly identified organism. The paper instead examines a broader propagation process involving AI systems, AI-generated information, culturally transmitted methods of using AI, commercial incentives, institutional adoption, and changing human behavior.

Just as important, the authors explicitly state that the viral analogy does not mean that every human–AI relationship is parasitic. Viruses in nature can be harmful, neutral, mutually beneficial, or even important contributors to evolutionary development. Human–AI coupling can likewise produce very different outcomes.

The central question is whether the relationship expands the person’s enduring capabilities or progressively substitutes for them.

Three States of Human–AI Coupling

The paper divides the human population into three theoretical categories: uncoupled or weakly coupled users, autonomous coupled users, and persistently dependent users. These are not clinical diagnoses. They are simplified categories used to construct a population-level mathematical model.

Uncoupled or Weakly Coupled Users

The first category consists of people who make little or no use of large-language models. The paper represents this group with the letter U.

Their cognitive environment may still include books, teachers, colleagues, calculators, search engines, computers, libraries, databases, and the Internet. No modern person is cognitively independent of all external tools. Writing itself is a technology for extending memory and organizing thought.

What distinguishes this category is that an LLM has not become a significant part of the person’s cognitive process. The individual continues to read, analyze, remember, compose, calculate, evaluate, and solve problems without routinely delegating those activities to generative AI.

This state preserves unaided cognitive practice, but it is not automatically the ideal condition. Refusing to use AI may protect existing abilities while preventing the person from acquiring important new capabilities. Someone who rejects calculators, computers, or the Internet entirely may retain certain manual skills while becoming less capable within a technologically advanced society.

The objective should therefore not be permanent isolation from AI. It should be the preservation of human cognitive autonomy while gaining the advantages AI can provide.

Autonomous Coupled Users

The second category consists of people who use LLMs while retaining their ability to reason, read, write, verify, and make decisions independently. The paper represents this group with the letter C.

These users are cognitively coupled to AI, but they are not controlled by that coupling. They treat the model as an additional component within a diverse cognitive environment rather than as their exclusive source of information or judgment.

An autonomous coupled user may ask an AI system to identify alternative explanations, challenge an argument, summarize technical literature, find gaps in a plan, generate possible solutions, or improve the organization of a document. The user still determines the objective, supplies relevant context, evaluates the alternatives, verifies important claims, and accepts responsibility for the result.

This is the most productive form of human–AI interaction. The AI is neither ignored nor blindly obeyed. It becomes a cognitive collaborator.

The user can ask: What have I overlooked? What assumptions am I making? What evidence contradicts my position? How could this design fail? What alternative strategies should I consider? How can this idea be explained more clearly? What knowledge from another discipline can be synthesized with this problem?

In that relationship, AI expands the number of ideas that can be explored, reduces the cost of testing alternatives, and accelerates the conversion of an initial concept into useful work. It can expand the Impact Radius of a Positive Trigger Event, accelerate its Diffusion, reveal opportunities across its Emergence Horizon, and support the Synthesis through which separate discoveries and ideas are combined.

That process can contribute to Compounding Advancement. Each improved capability, new insight, better process, and verified result becomes part of the enlarged capability base from which the next advancement can be produced.

The crucial test is what remains when the AI is removed. If working with AI helps a student understand a concept more deeply, formulate better questions, recognize weak evidence, and solve the next problem more effectively, the system has served as a scaffold and amplifier. If the student can produce an answer only while the AI is present, the system has become a substitute.

Persistently Dependent Users

The third category consists of users for whom AI increasingly replaces the cognitive operations needed to sustain independent competence. The paper represents this group with the letter D.

A dependent user may routinely accept AI-generated conclusions without understanding how they were reached. Writing, analysis, memory, calculation, planning, and decision-making are progressively delegated to the model. Alternative sources of information are consulted less often. Verification declines because it requires time and effort. The apparent convenience of AI reduces the friction that would otherwise force the person to think.

The immediate output may improve even while the user’s underlying competence declines. A person may produce a polished report without being able to explain its central argument. A student may submit correct computer code without understanding its logic. A manager may receive a sophisticated strategy without recognizing its unsupported assumptions. A user may remember less because the system can always be asked again.

The danger is not simply that the AI might produce an incorrect answer. The deeper risk is the gradual loss of the human ability to recognize that the answer is incorrect.

Dependence can become self-reinforcing. As people use certain abilities less frequently, those abilities weaken. As the abilities weaken, reliance on AI becomes more necessary. If schools and workplaces also reduce the expectation that people reason, write, calculate, or verify independently, the surrounding culture may stop reinforcing those abilities.

The individual then faces both personal and institutional pressure toward greater dependence.

The Mathematics of Adoption, Dependence, and Lock-In

The authors model movement among the three categories using a system of differential equations. People move from the uncoupled state to regular AI use through social, institutional, and platform-mediated exposure. Regular users may abandon AI and return to weak coupling, or they may progress toward persistent dependence. Dependent users may recover a more autonomous relationship through training, verification practices, deliberate cognitive effort, and continued access to alternatives.

The model includes five principal parameters.

The parameter lambda represents the effective rate at which LLM practices spread. Rho represents movement from regular AI use back toward uncoupled or weakly coupled cognition. Mu represents progression from autonomous use into persistent dependence. Sigma represents recovery from dependence to autonomous use. Kappa represents the collective social reinforcement of independent cognition.

The model’s most important feature is its ability to produce tipping points. If autonomous reasoning is strongly supported by schools, workplaces, families, and cultural expectations, society can resist dependency-producing forms of AI use. But if the proportion of people practicing independent cognition falls far enough, that protective environment may weaken.

Beyond a critical threshold, a relatively small increase in AI adoption can theoretically produce a much larger movement toward cognitive offloading and dependence.

The model also produces hysteresis: the conditions required to recover may be more demanding than the conditions that would have prevented the transition. Once AI-dependent processes become embedded in educational systems, workplaces, software platforms, and cultural expectations, merely reducing AI use to an earlier level may not restore the lost human capabilities.

This is technological lock-in applied to cognition.

The model demonstrates that such a transition is mathematically possible under its assumptions. It does not prove that this transition is occurring across society or establish the actual numerical thresholds. The authors do not present a longitudinal population study measuring movement among these three states.

The paper’s illustrative measure of independent cognitive competence assigns a value of 1.0 to uncoupled users, 0.5 to autonomous coupled users, and 0.1 to dependent users. Those numbers are modeling assumptions, not empirically measured facts. The authors acknowledge that properly scaffolded or augmentative AI use could preserve or increase subsequent independent competence.

The mathematical model should therefore be understood as a warning about a plausible systemic risk, not as proof of an inevitable cognitive epidemic.

Cognitive Immunization

The paper introduces “cognitive immunization” as a way to prevent beneficial AI use from becoming dependency-producing cognitive substitution.

Cognitive immunization does not mean banning AI, minimizing all AI use, or isolating children from modern technology. It means developing the knowledge, habits, institutional practices, and technological safeguards that allow people to use AI extensively without surrendering their cognitive autonomy.

The first component is maintaining independent competence. Students should continue reading complete works, writing without automated generation, performing appropriate calculations, constructing arguments, and solving selected problems without AI assistance. These exercises preserve the cognitive abilities upon which effective AI collaboration depends.

The second component is verification. Students should learn that fluent language is not evidence of truth. They should trace important claims to reliable sources, distinguish facts from interpretations, examine contradictory evidence, reproduce calculations, test computer code, and identify uncertainty.

The third component is metacognition—thinking about one’s own thinking. Students should be taught to ask which part of a task they understand, which part they have delegated, what assumptions the AI introduced, and whether they could explain or reproduce the result independently.

The fourth component is cognitive friction. Friction is not always inefficiency. Some productive difficulty is necessary for learning. Recalling information, struggling with a problem, revising an argument, and correcting an error help create durable knowledge. AI systems should not automatically remove every moment of difficulty before the learner has attempted the work.

The fifth component is reversibility. People should retain the ability to complete important tasks when AI is unavailable, compromised, incorrect, manipulated, or inappropriate. Schools and organizations should preserve alternative information sources, manual procedures, independent expertise, and opportunities for unaided practice.

The sixth component is diversity of cognitive resources. An LLM should not become the exclusive interface between a person and the world’s knowledge. Books, original documents, experiments, teachers, mentors, colleagues, direct observation, professional experience, and competing models remain essential.

These measures reduce progression toward dependence while increasing the possibility of recovery if dependence begins to develop.

Teaching Children to Think First and Collaborate Second

Children should be taught to use AI, but the sequence matters.

They should first develop a basic understanding of the subject and make an independent attempt. They can then use AI to critique, extend, test, or compare their thinking. Finally, they should evaluate the AI’s contribution, correct its mistakes, and produce a result they genuinely understand.

A practical educational principle is: think first, collaborate second, verify third, and explain the result independently.

Children should learn to use AI as a Socratic partner rather than an answer dispenser. Instead of asking the model to complete an assignment, they can ask it to question their assumptions, present counterarguments, explain why an attempted solution failed, or create a new problem that tests the same underlying concept.

They should sometimes be prohibited from using AI, particularly while foundational skills are being formed. At other times, they should be expected to use it because knowing how to collaborate with artificial intelligence will itself be a necessary capability.

Assignments should distinguish among unaided work, AI-assisted work, and fully collaborative work. Students should disclose how AI was used and identify which parts of the final product represent their own reasoning. Oral questioning can establish whether the student understands and can defend the submitted work.

Assessment should increasingly measure the student’s ability to define a problem, formulate a useful question, evaluate evidence, challenge an AI response, detect errors, synthesize multiple sources, and make a defensible judgment. These are the capabilities that become more valuable—not less valuable—when inexpensive generated content is everywhere.

The objective is not to preserve every traditional task exactly as it existed before AI. Innovation necessarily changes how work is performed. The objective is to ensure that each generation develops the intellectual foundation needed to direct the new technology rather than merely depend upon it.

From Cognitive Virus to Cognitive Collaborator

The paper’s most important contribution is not its frightening title. It is the distinction between complementary and substitutive technologies.

A substitutive AI performs the cognitive operation so completely that the user gradually loses the ability or incentive to perform it. A complementary AI improves immediate performance while helping the user maintain or increase enduring competence.

Used carelessly, an LLM can become a cognitive virus for a particular person: it can enter the individual’s working processes, encourage repeated delegation, displace independent practice, and create a self-reinforcing dependency.

Used deliberately, the same technology can become a cognitive collaborator and capability amplifier. It can expand access to knowledge, expose the user to alternative viewpoints, accelerate experimentation, strengthen analysis, connect previously separated ideas, and help transform imagination into innovation.

This is consistent with the central argument of The Imperative to Innovate. Technology does not determine its consequences independently of human purpose, institutional design, and responsible use. A new capability begins as a Positive Trigger Event, but its ultimate Impact Radius and Emergence Horizon depend upon how people direct its Diffusion and combine it through Synthesis with existing knowledge and human capabilities.

When AI strengthens the human ability to learn, reason, create, verify, and innovate, its contribution can become part of Compounding Advancement. The results of today’s collaboration become inputs to tomorrow’s discovery. Better tools support better processes; better processes generate better knowledge; and better knowledge enlarges the base from which the next advancement can occur.

But compounding can operate in the opposite direction. Unexamined dependence can weaken competence, weakened competence can increase dependence, and institutional adoption can spread that dependence throughout society.

The choice is therefore not between accepting AI and rejecting it. The choice is between designing human–AI relationships that cultivate capability and allowing convenience to displace competence.

We should teach children—and ourselves—to remain the source of purpose, the exercise of judgment, and the bearer of responsibility. Artificial intelligence should help us think more broadly, test ideas more rapidly, and create more effectively. It should amplify human intelligence without replacing the human cognitive foundations upon which meaningful innovation depends.

The proper response to the possibility of a cognitive virus is not technological quarantine. It is cognitive immunization—and the deliberate transformation of AI into a collaborator in the continuing advancement of human capability.