The paper's question is not about the model's intelligence but about the human-tool relationship. If we keep offloading memory, writing, and decisions to a language model, tool use can shift from a temporary choice to a habit and then to persistent dependence. Using a population model borrowed from epidemiology, the authors show this transition can be abrupt: past a threshold, a small rise in adoption triggers a rapid jump in persistent dependence.
The value of the paper is in moving the question: from what the model can do to human cognitive autonomy. The boundary of agency is not only inside the person or inside the model; it depends on how cognitive work is divided between human and tool. How much independent capacity to evaluate and reconstruct results we retain is the epistemic question of everyday use.
But this is a theoretical model, not an observation of real decline. The paper offers no causal evidence that language-model use has weakened real users' cognition, and the virus metaphor misleads if read literally. For tool designers, the more practical question is: what friction in the interface raises reversibility without destroying the model's usefulness?