--- title: Large-Language Models as a Cognitive Virus url: https://codereps.ai/blog/large-language-models-cognitive-virus published: 2026-10-28T16:41:38Z author: Dr. Nursultan Kabylkas tags: [llms, cognitive-offloading, learning, research, cal-newport, teaching] cover: https://codereps.ai/blog/images/f89ae2f0911f7b2db13cbf49fddbf6c4.png summary: A team of mathematicians modeled AI adoption like the spread of a virus. Their model shows tipping points and lock-in toward AI dependency, and points to a remedy: keep practicing the skills AI can replace. --- # Large-Language Models as a Cognitive Virus We have all heard the argument that AI might make us lazy thinkers. Why spend an hour solving a problem when ChatGPT can solve it in seconds? Why read a long article when AI can summarize it in a paragraph? But can we actually study this problem mathematically? I recently watched [an interesting discussion](https://podcasts.apple.com/us/podcast/an-unlikely-team-tackles-ai-brain-rot/id1515786216?i=1000795928156) by computer scientist Cal Newport on his *Deep Questions* podcast about a new research paper titled [*Large-Language Models as a Cognitive Virus*](https://arxiv.org/abs/2609.03344), published in September 2026. What makes this paper interesting is who wrote it. The authors are not primarily AI engineers or education researchers. They are a group of scientists working in areas such as applied mathematics, biophysics, and complex systems. Instead of asking whether AI makes an individual smarter or less intelligent, they ask a different question: **What happens to society when more and more people start relying on AI to think for them?** And to answer it, they borrowed an idea from epidemiology. ## 1. What do AI and viruses have in common? Remember when COVID-19 first appeared? Scientists were trying to predict how quickly the virus would spread, how many people would become infected, and what might happen if governments introduced social distancing. One way to study this was with mathematical models. A simple epidemic model divides the population into groups. For example: - **Susceptible:** People who haven't been infected. - **Infected:** People who currently have the virus. - **Recovered:** People who have recovered. People move between these groups over time. The model uses equations to estimate how quickly those transitions happen. For example, how many susceptible people become infected each day? Researchers can then change certain parameters, such as the rate of infection, to see how the simulated epidemic develops. Of course, an AI model is not a biological virus. But the researchers noticed an interesting similarity. AI usage also spreads through social interactions. Your classmates use ChatGPT, your colleagues start using it, your company introduces AI tools, and suddenly you are expected to use them too. And just as a virus can change the condition of its host, relying on AI might change the way we perform cognitive tasks. So the researchers asked: **Could we use similar mathematical equations to understand how AI adoption and dependency spread through society?** ## 2. Three types of people The researchers divided the population into three groups. **U — Uncoupled users** People who don't use AI, or use it very little. They rely on their own thinking and traditional tools. **C — Coupled but autonomous users** People who regularly use AI but can still think independently. They can write, reason, solve problems, and verify information without AI assistance. **D — Dependent users** People who rely heavily on AI to perform cognitive tasks. They increasingly delegate activities such as reasoning, writing, and problem-solving to the technology. The researchers then created three differential equations describing how people move between these groups. You don't need to understand differential equations to appreciate the idea. Imagine we start with a society where 90% of people don't use AI, 9% use it while thinking independently, and 1% are dependent. ``` Initially: U = 90% C = 9% D = 1% ``` Over time, more people discover AI and start using it. Some remain independent thinkers. Others gradually become dependent. The equations calculate how the percentages change. What makes this interesting is that the researchers can change the conditions of their simulated society and observe what happens. For example, what if the pressure to adopt AI increases? What if people are encouraged to practice thinking independently? What if dependent users can recover their skills? These conditions are represented by different parameters in the equations. ## 3. The surprising result: a tipping point One of the most interesting results is that AI adoption does not always increase smoothly. The model shows that, under certain assumptions, a small change in the pressure to adopt AI can cause a surprisingly large shift. Imagine a society where most people are not using AI. As AI becomes increasingly useful, more people begin adopting it. At first, the transition might be relatively slow. But eventually, the system can reach a critical threshold. Beyond that point, the simulated population abruptly moves toward widespread AI usage, including a substantial number of dependent users. The researchers call this a *tipping point*. Another interesting result is something called **hysteresis**. This is a complicated word for a fairly simple idea: returning to the previous conditions does not necessarily return the system to its previous state. Imagine that AI becomes so integrated into schools and workplaces that almost everyone uses it. Now suppose we reduce the pressure to use AI. Would society automatically return to the earlier situation? Not necessarily. According to the model, once widespread AI use becomes established, reducing adoption pressure to its previous level may not be enough to reverse it. A much larger change may be required. This is sometimes called technological lock-in. An important clarification: **the researchers did not discover an actual tipping point in human society.** They demonstrated that such a tipping point can exist in their mathematical model. Whether real-world AI adoption follows the same dynamics still needs to be investigated. ## 4. So how do we prevent cognitive dependency? This is where the paper becomes particularly interesting. The researchers explore which changes to their model could prevent society from ending up with too many people in the dependent group. There are three broad approaches. ### Approach 1: Reduce the pressure to use AI The most obvious approach is to discourage AI adoption. If people are less likely to start using AI, fewer people can become dependent on it. This may work in certain environments. For example, a school might prohibit AI during specific assignments or examinations. But avoiding AI altogether is probably unrealistic. The technology is already useful and is becoming integrated into everyday work. ### Approach 2: Make it easier to return to independent thinking Another approach is to encourage people to regularly perform tasks without AI. For example, someone who normally writes with ChatGPT might deliberately write some things independently. A student who uses AI to study mathematics might occasionally solve problems without assistance. The model suggests that making this return to independent activity easier can help prevent technological lock-in. It also shows that social environments matter. If schools, workplaces, and communities value independent thinking, they can help reinforce it. ### Approach 3: Prevent AI users from becoming dependent The third approach might be the most practical. Instead of trying to prevent people from using AI, we can try to ensure that they remain capable of thinking independently. Remember the three groups? **U → C → D** The goal is not necessarily to keep everyone in U. It is to allow people to move into C, where they benefit from AI, while reducing the probability that they move into D. And if someone does become dependent, the goal is to help them return to C. The researchers discuss activities such as independent problem-solving, verification of AI-generated information, metacognitive training, and periodic practice without AI. In the model, these interventions reduce dependency even when overall AI adoption remains high. This leads to an important conclusion: **The problem may not be how much we use AI, but whether we continue exercising the abilities that AI can replace.** ## 5. AI should help us learn, not replace learning There is an important distinction in the paper between two ways technology can assist human thinking. The first is called *scaffolding*. Imagine you are learning programming. You attempt to solve a problem, get stuck, and ask ChatGPT to explain a concept. You understand the explanation and then solve the problem yourself. In this case, AI helps you develop your skills. The second is called *substitution*. Imagine you receive the same programming problem, paste it into ChatGPT, copy the generated solution, and submit it. The assignment is completed, but have you learned to solve the problem? These two interactions might produce exactly the same output, yet lead to very different learning experiences. The paper discusses experimental research suggesting that unrestricted AI assistance can sometimes improve immediate performance while weakening subsequent independent performance. One well-known example is [a field experiment with nearly a thousand high school math students](https://doi.org/10.1073/pnas.2422633122): those given unrestricted GPT-4 did better on practice problems, but once access was removed they performed worse than students who never had it. It is important not to generalize this to every student or every AI interaction. The effects depend on how the tools are used. But it reinforces a simple idea: **being able to complete a task with AI is not the same as being able to complete it yourself.** ## 6. Perhaps we need a little time without AI I think the most useful takeaway is that we don't necessarily have to choose between embracing AI and rejecting it. We can do both things that matter: use AI to become more productive and deliberately preserve time to exercise our own minds. Consider physical fitness. Cars allow us to travel without walking long distances. Elevators allow us to reach the tenth floor without climbing stairs. We don't reject these technologies because they reduce physical effort. Instead, we recognize that our bodies still need exercise, so we make time for it. Perhaps cognitive fitness should work similarly. We can use AI for work, research, programming, and learning. But we should also create situations where we must think, reason, remember, write, and solve problems ourselves. Just as we don't expect riding in a car to improve our physical fitness, we shouldn't automatically expect completing tasks with AI to improve our cognitive abilities. And this is especially important in education. ## 7. What this means for CodeReps This research resonated with me because of something I observed while teaching programming. After ChatGPT became widely available, some students started submitting excellent programming assignments but struggled during supervised examinations where they had to write code independently. That observation was one of the motivations behind [CodeReps](https://codereps.ai). The idea is simple: give students a known pool of programming problems to practice, allow them time to prepare, and then assess their ability to solve a randomly selected subset independently. Students can benefit from AI while learning and preparing. But they also need time to practice without it, because ultimately they must demonstrate that they have developed the skill themselves. In a way, this is one practical example of what the paper calls *cognitive immunization*. CodeReps is not a scientifically validated treatment for AI dependency. That would require controlled studies and evidence about long-term learning outcomes. But the underlying educational philosophy is aligned with the researchers' recommendation: preserve opportunities for independent cognitive activity even as AI becomes more widely used. ## Final thoughts The paper does not prove that AI is making humanity less intelligent, nor does it predict that we will inevitably lose our ability to think. What it shows is that, under certain mathematical assumptions, the spread of AI dependency can create tipping points and make recovery more difficult. It also suggests possible ways to prevent that outcome. My favorite conclusion is perhaps the simplest one: **Keep using AI. Just don't stop practicing the things you want to remain good at.** Maybe the future of education is not about banning AI or allowing it everywhere, but about understanding when assistance helps learning and when it replaces the very effort that learning requires. And perhaps preserving independent thinking will become one of the most important responsibilities of educational institutions in the AI era. --- **Further reading** - Solé, R., Ruffini, G., Castaldo, F., et al. (2026). [*Large-Language Models as a Cognitive Virus*](https://arxiv.org/abs/2609.03344). arXiv:2609.03344. - Cal Newport, [*An Unlikely Team Tackles AI Brain Rot*](https://podcasts.apple.com/us/podcast/an-unlikely-team-tackles-ai-brain-rot/id1515786216?i=1000795928156), *Deep Questions* podcast, episode 427, October 12, 2026. - Bastani, H., Bastani, O., Sungu, A., et al. (2025). [*Generative AI without guardrails can harm learning: Evidence from high school mathematics*](https://doi.org/10.1073/pnas.2422633122). PNAS, 122(26).