A recent paper proposes a provocative way of thinking about the spread of large language models: as a “cognitive virus.” The metaphor is not literal. The authors are interested in what happens when AI becomes embedded in the way people think, work and communicate. If people increasingly delegate writing, remembering, reasoning and problem-solving to machines, the concern is that dependence could become self-reinforcing. Their theoretical model explores the possibility that LLM adoption could eventually reach a point where cognitive dependence becomes difficult to reverse. It also considers ways in which such dependence might be resisted or reversed.
It is an interesting warning, particularly because AI differs from most earlier technologies in the range of cognitive functions it can perform. A calculator can take over arithmetic; a search engine can take over information retrieval. A language model can participate in writing, analysis, explanation, brainstorming, argument and decision-making. The possibility that such a system could change the cognitive habits of its users deserves serious attention.
But there is another possibility hidden inside the same phenomenon. If AI becomes embedded in human cognition, does that necessarily mean that humans will surrender cognition to the machine? Could the opposite happen? Could the machine allow humans to explore considerably more ideas than they could explore unaided?
I would call that Cognitive Explosion.
1. Humans Have Always Re-adjusted Their Cognition
Human cognition has never operated independently of its tools. Writing changed what people needed to remember. Books allowed knowledge to accumulate outside individual brains and across generations. Calculators reduced the practical importance of performing arithmetic manually. Search engines changed the relationship between knowledge and recall, while smartphones placed external information and communication almost permanently within reach.
We did not simply adopt these technologies. We gradually re-adjusted our cognitive behaviour around them.
That does not mean every adjustment is beneficial. A calculator can certainly make someone less practiced at arithmetic, just as a search engine can encourage shallow information retrieval. But it would be strange to conclude from this that externalising cognition is inherently harmful. The more important question is what the technology enables us to do with the capacity it frees.
AI makes this question more consequential because it is not merely extending memory or calculation. It is beginning to participate in the generation and transformation of ideas themselves.
2. The Hidden Scarcity: Compatible Minds
This becomes particularly interesting when we consider the problem faced by a sophisticated thinker. One of the biggest constraints on serious thinking is not always intelligence. Sometimes it is simply the absence of another person capable of engaging with the problem.
Finding someone interested in technology is easy. Finding someone who understands technology, economics, psychology and history well enough to explore a problem at their intersection is much harder. As the combination of expertise becomes more unusual, the pool of potential collaborators becomes smaller.
This creates an underappreciated form of intellectual scarcity. A person may have plenty of ideas but very few people with whom those ideas can be properly tested.
The problem becomes particularly acute for someone whose interests or expertise cross several domains. The economist may not understand the engineering problem. The engineer may not understand the historical context. The historian may not be interested in the technological implications. The person who understands all three may be thousands of kilometres away, unavailable, or simply uninterested in having the conversation.
For much of human history, the solution was straightforward: find another human being.
The difficulty was that the right human being was often very difficult to find.
3. The Grandmaster Problem
Chess provides an unusually clear analogy. Finding someone who can play chess is easy. Finding someone who can genuinely challenge a grandmaster is not. At the highest levels, the number of suitable human opponents becomes extremely small.
Computer chess changed the economics of that problem. A system such as Stockfish can examine enormous numbers of possible continuations and evaluate them at a scale that is impossible for an unaided human player. Its significance is therefore not merely that it can announce the best move. It changes the practical search space available to the player.
That distinction is important when thinking about AI and intellectual work. The most interesting possibility may not be that an AI can give a person the answer to a difficult question. It may be that it allows the person to examine possibilities that would otherwise never have entered their thinking.
The machine does not necessarily make the thinker smarter. It can make the space available for exploration much larger.
4. From Chess Search to Idea Search
Consider what normally happens when someone has an interesting idea. They develop the idea using their existing knowledge and experience, think of several objections, look for examples, perhaps discuss it with another person and eventually settle on an interpretation. There is a natural limit to this process. Human attention is finite, memory is imperfect, time is limited and our existing knowledge strongly influences which possibilities occur to us in the first place.
AI can change that process. A thinker can ask for alternative explanations, then ask which assumptions those explanations depend upon. They can ask for the strongest objections, followed by possible responses to those objections. They can approach the same question from economics, history, psychology, engineering or evolutionary theory. They can ask the machine to deliberately misunderstand the argument, reverse its assumptions, or search for an apparently unrelated field containing a similar structure.
Much of what comes back will be mediocre. Some of it will be wrong. Some will merely rearrange familiar ideas. That is not a problem unique to AI; human brainstorming produces plenty of bad ideas too.
The important change is that the cost of generating another branch of thought becomes extremely low.
This is where the idea of Cognitive Explosion begins to make sense. The machine does not have to solve the problem. It can simply make it much cheaper to explore the problem from more directions.
5. AI as a Brainstorming Multiplier
Traditional brainstorming depends on the number and diversity of people available. Three intelligent people can usually produce more perspectives than one. Ten may produce more than three. But people have limited time, limited expertise and limited availability, and their knowledge inevitably overlaps.
AI changes the economics of another intellectual iteration. A person can ask for a different interpretation, challenge the response, introduce new information and continue the conversation. They can explore a branch of thought for ten minutes and abandon it without having consumed another person’s afternoon.
That does not guarantee better ideas. It does something more basic: it makes more exploration economically possible.
This is particularly valuable when the objective is not to obtain an answer but to discover what the answer might even look like.
6. Substitution Versus Amplification
There are obviously circumstances in which AI substitutes for human cognition. Someone can ask a model to write an article, summarise a book or solve a problem and simply accept the result. If that pattern becomes habitual, the cognitive-virus concern becomes increasingly relevant. The user is no longer using the machine to extend their thinking; the machine is performing the thinking they would otherwise have performed.
But there is another way to use exactly the same system. A person can bring an idea to AI precisely because they want it challenged. They can ask for weaknesses, counterarguments, alternative explanations and connections to other fields, and then use their own knowledge and judgment to decide which responses are valuable.
That is not cognitive substitution in the same sense. It is cognitive amplification.
The distinction is not in the model. It is in the relationship between the human and the model.
7. An Irreverent Update to an Old Saying
There is an old saying that great minds discuss ideas, average minds discuss events and small minds discuss people. It is often attributed to Eleanor Roosevelt, although the attribution is uncertain.
In the age of WhatsApp and ChatGPT, one could make a deliberately irreverent update: “Small minds only chat. Average minds do WhatsApp chat. Great minds discuss ideas with ChatGPT.”
The joke works because there is a serious observation underneath it. Intellectual companionship has always been scarce. If someone wanted to test an unusual idea, they needed another person capable of understanding it. If the idea crossed several domains, finding that person became more difficult. If it was highly specialised, the pool became smaller still.
AI does not eliminate this problem, and it certainly does not replace genuine expertise. But it creates something new: on-demand intellectual interaction.
A thinker no longer necessarily has to wait for the right person to become available before exploring an idea.
8. The Cognitive-Virus Argument Is Only Half the Story
This brings us back to the cognitive-virus hypothesis. The paper is concerned with a trajectory in which AI becomes increasingly embedded in human cognition through dependence and cognitive offloading. That is a plausible trajectory and deserves investigation.
But embedding AI into cognition does not logically imply that cognition must become weaker.
There is another trajectory in which AI becomes embedded because humans discover that it allows them to explore ideas more extensively. The machine generates possibilities, exposes weaknesses, introduces unfamiliar perspectives and helps the thinker move across disciplinary boundaries.
The same technology could therefore produce very different cognitive outcomes depending on how it is used.
This is why I think the broader idea of natural cognitive re-adjustment is useful. Humans adapt to their cognitive environment. The interesting question is not simply whether AI changes that environment, but whether the resulting adaptation reduces human agency or expands it.
9. When Ideas Stop Being Scarce
There is an even more interesting consequence. If AI makes the generation of possibilities extremely cheap, ideas themselves may cease to be the primary bottleneck.
Judgment becomes the bottleneck.
A person who can generate five hypotheses has one problem. A person who can generate five hundred has another. The second person does not necessarily need more imagination. They need better mechanisms for deciding which possibilities deserve attention.
They need to recognise when an apparently novel idea is actually nonsense. They need enough domain knowledge to detect confident errors. They need to distinguish an interesting connection from a superficial one. They need to know which questions deserve another hour of exploration and which should be abandoned after two minutes.
AI may therefore increase the value of human judgment even while reducing the amount of routine cognitive labour humans perform.
That is a much more interesting possibility than the simple idea that AI either makes us smarter or makes us dumber.
10. What Stockfish Really Teaches Us
The important lesson from Stockfish is not simply that computers calculate faster than people. It is that a machine can radically expand the number of possibilities a human can practically investigate.
The grandmaster still needs to understand the position. The engine does not eliminate judgment. But it allows the player to examine lines that would otherwise remain beyond practical human calculation.
AI may be approaching an equivalent transition in intellectual work.
The machine can generate candidate ideas and arguments at enormous scale. The human can decide which ones are interesting, which ones are wrong, which ones deserve further investigation and which ones lead somewhere unexpected.
The machine therefore does not necessarily replace the thinker. It can enlarge the territory the thinker is capable of exploring.
11. Cognitive Explosion
This is the counterweight to the cognitive-virus argument.
The virus metaphor highlights the possibility that humans will progressively surrender cognitive functions to machines. The cognitive-explosion metaphor highlights the possibility that machines will make it possible for humans to explore vastly more cognitive possibilities than they could explore alone.
These are not mutually exclusive futures. A person could use AI to avoid thinking about routine matters while simultaneously using it to think more deeply about difficult ones. Someone could become dependent on AI for basic writing while becoming considerably more ambitious in the questions they are able to investigate.
The decisive issue may therefore not be whether AI is good or bad for cognition. It may be what humans ask it to do with their cognition.
If the machine becomes an oracle whose answers are accepted, the cognitive-virus concern becomes increasingly plausible. If it becomes a tool for generating possibilities, challenging assumptions and extending exploration, something quite different is happening.
The latter is what I mean by Cognitive Explosion.
12. The New Cognitive Environment
Perhaps the biggest mistake would be to imagine that AI represents a completely new category of influence on the human mind. Humans have always been shaped by their cognitive environment. Conversation, education, books, newspapers, television, the internet, social media and YouTube have all altered what people notice, remember, believe and talk about.
AI is different in degree and perhaps eventually in kind because it is interactive. It can respond to the individual rather than simply broadcasting information to a mass audience. It can follow a line of reasoning, generate alternatives and participate in an extended intellectual exchange.
That makes natural cognitive re-adjustment particularly important. We are likely to adapt to this new environment just as we adapted to previous ones. The outcome will depend partly on what habits we build around the technology.
The danger is that we use the enormous convenience of AI to eliminate the need for independent thought. The opportunity is that we use the enormous availability of AI to expand the amount of thought we can undertake.
Those are very different futures.
13. From Cognitive Virus to Cognitive Explosion
The cognitive-virus research gives us a valuable warning. It reminds us that a technology capable of performing cognitive work can also alter the behaviour of the people who use it. We should take dependence, skill erosion and loss of autonomy seriously.
But we should not stop there.
There is an equally important possibility that deserves investigation: AI may be less important as a replacement for individual acts of cognition than as an expansion of the search space within which human cognition operates.
That is why the Stockfish analogy matters.
Stockfish did not make chess larger. It made more of chess practically searchable.
AI may be doing something similar with ideas.
For the first time, an individual thinker can potentially have an always-available system with which to test an argument, explore an alternative, cross a disciplinary boundary, generate possibilities and continue a line of inquiry long after a normal human conversation would have ended.
The result may not be that humans think less. It may be that one human can explore vastly more possibilities than one human could ever have explored alone.
That is the cognitive explosion.
And it creates a new responsibility. Once the cost of generating ideas collapses, generating ideas is no longer enough. The scarce and valuable human capability may increasingly be the ability to recognise which ideas deserve to survive the explosion.
Source: Ricard Solé et al., “Large-Language Models as a Cognitive Virus”, arXiv, September 2026. The paper presents a theoretical model of LLM diffusion, dependence and possible cognitive lock-in; it does not by itself establish that a population-wide cognitive decline is occurring.