Cognitive offloading is the habit of handing a mental task to something outside the head. Writing a list instead of remembering it, using a calculator, checking a map, and now asking a language model to draft, summarise or decide. Psychologists have studied it since long before generative AI. Risko and Gilbert’s 2016 review set out the general pattern: people offload more as the aid becomes more reliable, the offloading is efficient, and the skill that is offloaded gets less practice.
It matters because generative AI can take the whole task rather than one step of it. A calculator removes the arithmetic and leaves the problem. A chatbot can remove the problem. The controlled studies that measured this, in schools, workplaces and clinics, found the same pattern: performance rises while the tool is present and falls below the unassisted baseline once it is removed, in some cases after a single ten-minute session.
The practical distinction is between offloading a step you have already mastered and offloading a step you have not. An expert who offloads recall can usually recover it. A student or junior who offloads the reasoning before acquiring it has nothing to recover, which is why tools that give hints rather than answers protect learning where answer-giving tools erode it.