Everyone working in technology right now is using AI. Not occasionally, not experimentally, but daily and at volume. Writing, researching, summarising, reasoning, generating, reviewing. The tools have become fluent enough that reaching for them feels natural, almost automatic. That is precisely what makes the question worth asking: what is it doing to us?
A handful of studies published over the last year suggest the answer is not straightforwardly positive. The concern is not that AI is bad or that using it makes you incompetent. The concern is subtler and, in some ways, more dangerous. It is that convenience, at sufficient scale, begins to replace the mental effort that builds capability in the first place.
What the MIT Research Found
Researchers at MIT Media Lab ran a study titled "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." They divided participants into three groups. One wrote essays using only their own thinking. A second used a search engine. A third used a large language model. The researchers measured brain activity using electroencephalography throughout the task.
The results were striking. Brain connectivity scaled in direct proportion to how much external support participants received, but in the wrong direction. The group relying solely on their own cognition showed the strongest and widest-ranging neural networks. The search engine group showed intermediate engagement. The LLM group showed the weakest overall coupling across brain regions associated with critical thinking and memory.
It was not just that LLM users thought less during the task. They also retained less afterwards, and reported a diminished sense of ownership over the work they had produced.
The study is small and has not yet been peer-reviewed. That is worth noting. But it fits into a broader pattern that is harder to dismiss than a single piece of research.
The Confidence Paradox
In early 2025, researchers from Microsoft Research and Carnegie Mellon University published a study through the ACM CHI conference examining how generative AI affects critical thinking in real workplace settings. They surveyed 319 knowledge workers across 936 actual uses of AI tools, asking participants to reflect on how much cognitive effort they applied to each task.
The headline numbers are revealing. In knowledge work tasks, 72 percent of participants reported applying less mental effort when using AI. In comprehension tasks, that figure rose to 79 percent. In synthesis, 76 percent. Even evaluation, which one might expect to demand independent judgment, showed a 55 percent reduction in reported effort.
The finding that stands out, though, is the confidence paradox. Workers who expressed high confidence in the AI tool engaged the least critical thinking of any group. Workers who expressed high confidence in their own abilities engaged the most. The implication is that trusting AI deeply is not a careful posture. It is an abdication of the verification behaviour that sound use of any tool requires.
The researchers also found that AI-assisted workers produced a narrower, less diverse set of outcomes for the same task compared to those working without AI. The tool was not expanding thinking. It was converging it.
Cognitive Offloading and What It Costs
Cognitive offloading is the practice of using external tools to reduce mental load. Writing things down, using a calculator, setting a reminder. These are all forms of offloading, and most of them are benign and sensible. The question is what happens when the offloading becomes habitual and covers cognitive tasks that were building capability.
A 2025 study examining LLM use under time constraints found that younger participants showed higher dependence on AI tools and significantly lower critical thinking scores than older cohorts who had developed those skills before AI assistance was readily available. A separate study found that participants who used ChatGPT for essay tasks showed meaningfully lower cognitive engagement scores than those who worked without it, a pattern the researchers described as producing more “lazy” thinking.
The mechanism is not mysterious. Skills that are not exercised atrophy. If reasoning through a problem from first principles is repeatedly replaced by asking a model to do it, the neural pathways that support that reasoning get less use. This is not unique to AI. The same concern applies to GPS navigation and spatial awareness, or to calculators and mental arithmetic. What is different with AI is the scope. It covers language, logic, analysis, synthesis, evaluation and creative generation, essentially the full range of knowledge work.
The Faster-Is-Not-Better Problem
One framing that comes up repeatedly in this conversation is that AI accelerates the routine parts of thinking, leaving more time for the harder, more valuable parts. That is a reasonable hypothesis. The evidence, so far, does not support it strongly.
The MIT study showed that LLM users did not use the time freed by AI assistance to think more deeply. They thought less overall. The Microsoft study showed that workers did not apply the effort saved on generation to more rigorous evaluation. They evaluated less carefully. The hypothesis assumes a discipline around how freed cognitive capacity gets redeployed, and most people, under normal working conditions, do not exercise that discipline. They move on to the next task.
There is also a compound effect that is worth naming. If AI-generated output looks clean and well-structured, it passes review more easily. If it passes review easily, less scrutiny gets applied to whether it is actually correct. Over time, the standard of what counts as done quietly lowers, and no single decision caused it.
AI as Amplifier, Not Leveller
The most honest way to think about these tools is as amplifiers rather than equalisers. A person with strong foundational knowledge, genuine critical thinking skills, and a disciplined approach to verification benefits from AI in the way a skilled craftsperson benefits from a better tool. The core capability is already there. The tool extends it.
A person who uses AI to substitute for foundational knowledge they have not developed, or who accepts outputs without understanding them, gets something different. The gap between what they appear to produce and what they actually understand widens. That gap becomes a liability the moment something goes wrong and someone needs to reason through it from the ground up.
The research does not argue that AI makes everyone worse. It argues that the direction of effect depends almost entirely on how the tool is used, and that the default mode of use, which is to accept and move on, trends toward cognitive atrophy rather than cognitive development.
The Education Dimension
The implications for professional development are particularly serious and not widely enough discussed. Skills in complex domains are built through struggle. Reading something difficult, not quite understanding it, sitting with the confusion, working through it, and eventually reaching clarity is how comprehension is actually formed. Receiving a clear summary from a language model that skips the struggle produces a feeling of understanding without the neural substrate that makes understanding durable and transferable.
If you are early in your career and you use AI to generate solutions you cannot explain, you are not building the knowledge those solutions represent. You are accumulating a vocabulary without a grammar. It will feel fine until you need to debug something at 2am that behaves in a way no model predicted, or until you are asked to extend a system you thought you understood.
The same applies to experienced practitioners who stop doing things the hard way. The judgment that comes from working through difficult problems is not a fixed asset. It requires maintenance.
What Responsible Use Actually Looks Like
None of this means stopping. These tools are genuinely powerful, and refusing to use them would be its own kind of professional negligence. The question is how to use them without paying the cognitive cost the research describes.
The distinction that matters most is between using AI to do your thinking and using AI to accelerate thinking you are actually doing. The former replaces cognitive work. The latter supports it. In practice, this means engaging with the model as a collaborator you interrogate rather than an oracle you accept. It means reading AI outputs critically, asking yourself what assumptions they rest on, what they might be missing, and whether the reasoning holds. It means being willing to work through problems without AI assistance often enough to keep the underlying skills in shape.
There is also a legitimate case for deliberate friction. If you are learning something new, refusing to ask AI for the answer until you have genuinely tried to reason through it yourself is not inefficiency. It is investment. The time spent in confusion before clarity is where the actual learning happens.
The researchers who ran the MIT study put it simply: the participants who let AI do the writing did not just produce less original work. They remembered less of what they had written and felt less connection to it. Convenience extracted a price from the very thing the work was supposed to develop.
A Question Worth Sitting With
The question in the title of this post is not rhetorical, and the honest answer is: it depends. AI, used well, does not make you dumber. It can extend capability, accelerate learning, and free up cognitive resources for work that genuinely requires them. But AI used as a substitute for thinking, accepted without scrutiny, relied on at the expense of developing foundational knowledge, does appear to erode the cognitive skills that make someone capable of doing hard work independently.
The research is early. But the mechanism is not speculative. If you stop exercising a capability, it weakens. The same principle that governs physical fitness governs cognitive fitness. The question is not whether you use AI. It is whether you are still doing the reps.