Researchers Describe LLMs as a ‘Cognitive Virus’ That Encourages Mental Offloading

# What the paper says An international team of researchers lays out a theoretical framework that treats large language models (LLMs) as a kind of cognitive virus. The phrase is meant to capture two linked points: LLMs spread rapidly through social channels, and their use can encourage people to offload cognitive work in ways that lower unaided competence.

LLMs are already common across social media and everyday services, the paper notes. That ubiquity makes it easy for people to adopt model-assisted habits because they see others doing it and because the tools are useful. The researchers describe that spread as contagion-like: adoption occurs through observation, imitation, and social learning.

# How cognitive offloading works Humans have long externalized parts of cognition into tools such as writing, language, and institutions. LLMs represent a further step: external cognitive machinery that appears agent-like and actively completes mental tasks for us.

When people rely on models to generate answers, draft text, or check reasoning, they can stop practicing core skills: independent problem solving, source verification, critical discussion, and step-by-step reasoning. The paper highlights classrooms as an area where offloading is already visible: students use LLMs to do work that previously required active thinking.

Because the behavior is social, loss of autonomy can become self-reinforcing. If many people stop exercising certain skills, those skills at the population level can decline quickly.

# Risks described The authors frame the risk in practical terms: reduced cognitive competence across communities, changes to how people form beliefs, and social effects that are only beginning to be understood. They stop short of claiming specific, quantified harms, offering instead a conceptual model that links tool adoption to diminished mental practice.

They explicitly name commonly used LLMs such as ChatGPT and Anthropic's Claude as examples of the kind of tools whose proliferation drives the dynamics they describe.

# Proposed protections: cognitive immunization The paper proposes a set of concrete protections it calls cognitive immunization. These are practices and institutional designs intended to keep human cognition active rather than outsourced entirely:

• Maintain practices of unaided problem solving as a regular requirement.
• Preserve verification habits: independently check facts and sources instead of relying solely on model output.
• Keep spaces for critical discussion where people have to articulate reasoning without model assistance.

• Protect and teach non-AI skills so they remain functional in daily life.
• Design education so LLMs assist parts of a task rather than completing the entire task for students.

# Why this matters now LLMs are already embedded in many parts of daily life. The paper argues that unless individuals and institutions adopt measures that sustain cognitive practice, offloading may shift collective competence. The authors present their framework as a way to identify where policy, pedagogy, and personal habits should intervene.

# Caveats The paper has not yet been peer reviewed. It is presented as a theoretical framework and set of proposals rather than a definitive empirical measure of harm. It draws on observable behaviors—classroom use, social spread, and model ubiquity—rather than new experimental data in this publication.