Copy, paste, forget: how AGI is making us stupid and how to stop it
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
When a student uses AI to generate an essay, copies it, and submits it without reading, something invisible happens: a small piece of their cognitive capacity quietly atrophies. This is not a moral story about cheating. It is a scientific finding about cognitive debt — the accumulated cost of outsourcing thinking to machines. As Artificial General Intelligence becomes embedded in education, workplaces, and daily life, the risk is not that machines will become smarter than us. It is that we will become less capable of thinking for ourselves.
The Evidence of Cognitive Atrophy
The research is unambiguous. A study from MIT Media Lab, led by neuroscientist Nataliya Kosmyna, found that participants using ChatGPT experienced a 47% drop in neural connectivity — from 79 functional brain connections to just 42. More troubling: over 83% of ChatGPT users were unable to recall key details of the essays they had just written, compared to 89% success in the self-writing group. The AI-assisted essays were judged as good or better than human-written ones, but the students who "wrote" them could not remember what they had said.
The study also found a 32% reduction in germane cognitive load — the kind of mental effort tied to understanding, learning, and reflection. Participants who wrote without AI often described a greater sense of satisfaction and engagement, evidence that slower, more effortful work still holds value.
The Cognitive Debt Cycle
A survey of STEM students across five North American institutions found that regular GenAI users reported significantly less cognitive engagement — less reflection, less need for understanding, less critical thinking . Researchers describe a cognitive debt cycle: just as technical debt accumulates when short-term convenience replaces principled design, cognitive debt accumulates when students repeatedly delegate cognitive work to AI. The danger is that GenAI may restructure the learning environment so that delegation appears rational, gradually training the brain to reassess the cost of effort.
The most sobering finding came when long-time ChatGPT users were asked to complete tasks without assistance. Their performance declined more than those who had never used AI at all. It was as if their brains had forgotten how to work independently.
The GPS Metaphor
The analogy to GPS navigation is precise. Habitual reliance on turn-by-turn directions has been shown to impair spatial memory and hippocampal engagement. AI functions as a "GPS for thinking," short-circuiting the metacognitive loop — the student's ability to plan, monitor, and evaluate their own thinking.
Research from the University of Zurich notes that learners often do not choose the most effective study strategies. They prefer re-reading and highlighting — pleasant but ineffective — over self-testing, which is more effortful but produces better learning. AI amplifies this tendency. When learners can access apparently complete answers with zero friction, desirable difficulty — the effort that produces learning — is bypassed entirely.
The Epistemic Confinement
A mixed-methods study of 188 graduate students at the University of Melbourne introduced a troubling concept: epistemic confinement. Students reported feeling they were thinking independently while actually operating entirely within the analytical boundaries constructed by AI. Over 95% of the cohort fell into an "efficiency trap," offloading foundational interpretation and analysis to AI while struggling to preserve higher-order reasoning to maintain a semblance of intellectual agency.
This is the mirror image of the MIT finding that AI users cannot remember what they wrote. The student is present. The thinking is not.
The Institutional Mimicry
The problem is not confined to individual study habits. At the institutional level, assessment systems themselves are mimicking rather than correcting this decline. When essays can be AI-generated, the traditional "end product" loses its validity as evidence of learning. As one education technologist observed, "AI did not break homework: it exposed what was already broken".
Professor Jeremy Rentz of Trine University proposes practical alternatives: frequent in-class exams, reclassifying homework as practice rather than a primary graded artifact, and embedding review sessions during class time. For out-of-class work, he recommends grading the process rather than the product.
The University of Zurich's guidance is more direct: there must be phases where students work independently without AI. Students must write their own summaries, generate their own questions, and answer them themselves before checking what they know. "Only when learners have sufficient prior knowledge can they correctly evaluate AI output".
Productive Friction: Designing Against the Frictionless
The most systematic response comes from the intersection of virtue epistemology and AI design. Researchers propose the concept of productive friction — not maximizing difficulty, but calibrating friction to retain enough resistance for judgment to develop while removing friction that serves no pedagogical purpose.
This framework distinguishes three types of friction design:
Selective friction: user-activated modes that impose verification, requiring cross-referencing before content can be copied or saved.
Obstructive friction: deliberate delay or interruption before AI output is accepted, creating space for reflection.
Protective friction: safeguards that prevent irreversible cognitive harm, such as blocking AI use during foundational learning phases.
The key insight is that friction is not the enemy of learning. It is its substrate. The challenge is to design AI systems that preserve the effort that produces understanding while removing the effort that produces only frustration.
A 2026 systematic review of 67 studies confirms this approach. ChatGPT worked best when educators designed recursive learning cycles: generate, critique, revise, reflect. It performed poorly when students used it passively as a shortcut for task completion. The strongest results appeared in courses where students had to critique AI outputs, compare alternatives, verify information, and justify decisions.
The Path Forward
For Global Future Nexus, the cognitive atrophy caused by frictionless AI is not an educational problem. It is a civilizational one. If the generation that builds AGI cannot think without it, the governance of AGI will be exercised by those least equipped to understand what they are governing.
The path forward requires three commitments.
First, cognitive friction by design: embedding productive difficulty into AI-assisted workflows, especially in education.
Second, process-based assessment: evaluating how students think, not just what they produce.
Third, independent thinking phases: mandating periods where AI is set aside entirely, so that the cognitive muscles needed to evaluate AI output are developed before they are needed.
The intelligence we build must not be built on the atrophy of our own. The copy-paste generation is not a lost cause. But it will be if we do not design the resistance that makes thinking necessary.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
