Is AI making us stupid? The question is provocative, but it bundles together several different concerns. Forgetting how to perform a skill because we no longer practise it is one thing. Losing basic capacities such as attention or working memory is something much larger.
In a recent article, Trent Cash and colleagues use this distinction to make sense of the evidence. Their argument fits directly within my broader coverage of AI and cognition: AI can weaken particular skills when it takes over the practice that builds and maintains them, while claims of widespread cognitive decline remain premature.
Summary of the Study
This is a short evidence-based commentary. It reports no new experiment and follows no systematic-review protocol. The authors draw on research from cognitive science, learning, medicine, and human-AI interaction to examine what cognitive offloading may do over time.
Cognitive offloading occurs when we transfer mental work to an external aid. Calculators, GPS, search engines, and written reminders already do this. Generative AI extends the practice because it can perform complex work involving explanation, writing, summarising, and problem-solving.
The authors separate learned skills from basic cognitive abilities. Skills such as solving equations, writing, or detecting medical abnormalities depend on knowledge and repeated practice. If AI repeatedly performs the relevant work, people may fail to acquire the skill or gradually lose it.
The evidence they discuss includes students who performed well with answer-generating AI during practice but struggled when it was removed, a pattern also examined in my review of AI without learning guardrails. They also describe professionals whose unaided performance declined after working with AI assistance. In both cases, the risk concerns a specific practised skill.
The case for damage to foundational cognitive abilities is much weaker. Working memory and selective attention are broad capacities that support many kinds of thinking, and research suggests they are relatively resistant to sweeping change.
This distinction is useful because public discussion often moves too quickly from “AI may weaken a skill” to “AI is making people less intelligent.” Current evidence does not justify that leap. At the same time, the absence of long-term evidence should not be treated as reassurance, especially for children who may begin offloading before foundational skills are secure.
How AI is used changes the outcome. A tool that supplies finished answers can remove the practice students need, while a tutor that asks questions, identifies gaps, or provides feedback may keep them cognitively involved. The authors’ argument closely matches the distinction between beneficial offloading and outsourcing discussed in The Cognitive Risk of AI.
Still, their article is selective and conceptual. It does not provide a formal search strategy or a systematic assessment of the evidence, and the long-term effects of routine AI use remain unknown. It gives us a clearer way to frame the problem, but it cannot deliver a final verdict on human cognition.
Practical Insights
The classroom implication is not that every use of AI should demand maximum effort. Offloading routine work can free attention for a more important goal. The question is whether AI is removing the mental activity students need to learn.
- Identify the skill before introducing AI. Ask what students should still be able to do when the tool is gone. If the answer is “write an argument,” “solve this kind of problem,” or “evaluate evidence,” students need regular unaided practice in that skill.
- Delay complete answers. Structure AI use so that students attempt the task first, request a hint or explanation, and return to their own solution. A short independent reconstruction after AI use can reveal whether understanding has actually developed.
- Alternate assisted and unassisted work. Students can use AI during one part of an activity and then complete a related task without it. This makes the difference between supported performance and independent learning visible.
- Ask students to name what they offloaded. A brief reflection might ask: What work did the AI perform? What thinking remained yours? Could you complete a similar task tomorrow without assistance? This can help counter the weak monitoring associated with metacognitive laziness.
- Protect early skill development. Novices have less knowledge for evaluating AI output and more foundational learning still to build. AI support should leave enough intellectual work for them to form the skill in the first place.
Researchers now need longitudinal studies that distinguish skill acquisition, skill retention, knowledge retrieval, metacognition, and basic cognitive abilities. Different ages, disciplines, levels of prior expertise, and forms of AI support should be studied separately. A broad question about whether AI makes us stupid is unlikely to produce a useful answer if these very different outcomes remain mixed together.
The answer, for now, is narrower than the headline. AI can leave us less capable of doing the work we repeatedly hand over to it. That is a genuine educational risk. It is also a design problem: the same technology can provide the finished performance or help learners build the capacity to perform independently.
Reference
Cash, T. N., Kelly, M. O., Macnamara, B. N., & Risko, E. F. (2026). Is AI making us stupid? Trends in Cognitive Sciences, 30(8), 673–675. https://doi.org/10.1016/j.tics.2026.06.004
