AI can be a shortcut around thinking or a tool for thinking harder. The difference depends on how we engage with it.
Every technological revolution produces its own anxieties about the human mind. Calculators supposedly weakened arithmetic, search engines threatened memory, and smartphones were accused of destroying attention spans. Now artificial intelligence faces a more unsettling charge: it is making us stupid.
The concern is understandable. When a chatbot can write an essay, summarise a book, solve a mathematical problem or formulate an argument within seconds, why expend the mental effort? Yet the increasingly popular claim that AI is eroding human intelligence runs considerably ahead of the scientific evidence.
What research currently suggests is more nuanced: AI can encourage intellectual dependency, but it can also enhance learning and reasoning. Much depends on whether we use it to replace thinking or support it.
A widely discussed 2025 study by researchers at Microsoft and Carnegie Mellon University surveyed 319 knowledge workers about their interactions with generative AI. It found that greater confidence in AI was associated with less reported critical-thinking effort. The findings raised legitimate concerns about users becoming passive recipients of machine-generated outputs.
But there is an important distinction between reduced cognitive effort and diminished cognitive ability. The study relied on self-reported experiences, not measurements of long-term intellectual decline. Spending less effort on a task does not necessarily mean losing the capacity to perform it.
Another investigation, Your Brain on ChatGPT, conducted by researchers associated with MIT Media Lab, attracted considerable attention. Using electroencephalography, researchers observed differences in neural connectivity among participants writing essays with ChatGPT, search engines or no external assistance. Those using ChatGPT generally exhibited weaker connectivity and less ownership of their writing.
The findings warrant attention, but not alarmist conclusions. The research involved a relatively small sample, focused on a specific writing exercise and was released as a preprint. Its authors explicitly cautioned against interpreting the results as evidence that AI makes people stupid or damages their brains.
Indeed, the broader evidence complicates the pessimistic narrative.
A 2025 randomised controlled study involving Harvard undergraduate physics students found that a carefully designed AI tutor produced substantially greater learning gains than conventional active-learning classroom instruction, in less time. Rather than eliminating intellectual engagement, the system helped students understand difficult concepts through personalised guidance.
Conversely, another 2025 study involving high-school mathematics students in Turkey found that unrestricted access to generative AI improved performance during practice but reduced subsequent performance when students had to solve problems independently. Crucially, an AI tutor designed to offer hints instead of ready-made answers largely avoided this negative effect.
Together, these findings reveal a distinction that much public commentary overlooks: the problem may lie less in artificial intelligence itself than in the intellectual behaviour it encourages.
Consider two students preparing for an examination. One asks AI to solve every problem and memorises the responses. The other attempts each problem independently, uses AI to identify mistakes, requests alternative explanations and challenges its reasoning. Both are using the same technology, but only one is consistently exercising the skills necessary for deeper learning.
The distinction extends beyond classrooms.
A journalist who unquestioningly reproduces an AI-generated analysis risks surrendering editorial judgment. A researcher who uses AI to examine competing hypotheses may sharpen it. A professional who delegates every decision may gradually become dependent on automated recommendations, while another who uses AI to interrogate assumptions may develop more sophisticated judgment.
None of this suggests that cognitive offloading is inherently harmful. Human civilisation has progressed by developing tools that reduce routine mental labour, from writing and libraries to spreadsheets and computers. The relevant question is what people do with the intellectual capacity those tools release.
Nor should we confuse fluency with understanding. Generative AI produces remarkably convincing language, including when its conclusions are wrong. Its greatest danger may therefore be less the disappearance of human intelligence than the creation of an illusion of knowledge: answers become easier to obtain while the discipline of evaluating them becomes easier to neglect.
Educational institutions and workplaces must consequently rethink how they assess intellectual capability. Producing polished outputs is no longer sufficient evidence of understanding. Explaining reasoning, defending conclusions, identifying errors and demonstrating independent competence matter more than ever.
Ultimately, asking whether AI makes us stupid may be the wrong question. The evidence is too limited to support such a sweeping verdict, and the technology is too versatile to produce a single intellectual outcome.
The more consequential question is whether we are cultivating habits of inquiry or habits of dependence.
AI will not determine the answer on its own. We will, through the ways we choose to use it.