Abstract

Generative AI has only been around for three years but is already a standard feature of the majority of student work in the UK. Undergraduate students’ usage has shot up to near-universal level: 66% to 95% between 2024 and 2026. This piece questions a line of thinking that has emerged faster than it should have: is giving thinking to a machine the same kind of cognitive offloading that human beings resort to time and again, or is this altogether a different thing? By drawing on Risko and Gilbert’s (2016) idea of offloading, they argue only works as long as the offloader’s selfassessment of their own ability remains accurate, the authors of this paper develop a tool. Through experiments with navigation, search and generative AI, we observed that the action of offloading did not change, but what is being transferred did change. Instead of handing over just part of the thought process, students end up handing over the very thing that monitored whether the thought was good or not. Hence, the danger is not the tool itself but when this happens, as the very delegation that enables a master to delegate, will limit the learner whose internal guide develops fully. Since Bastani et al. (2025) showed that by employing a well-designed tutor, every gain could be preserved without any loss happening, then this damage can be attributed to the way the tool is constructed. The essay recommends Commit-Before-Consult which is a minimal-risk way of tuning up and explains it.

Copyright © 2026 Scholar of Tomorrow. All SoT articles are distributed under the attribution non-commercial, with no derivative license. This means that anyone is free to share, copy, and distribute an unaltered article for non-commercial purposes, provided the original author and source are credited.