What the research actually shows

Start with the honest good news. In a Harvard Business School and Boston Consulting Group field experiment (Dell'Acqua and colleagues, 2023, later published in Organization Science), 758 consultants using GPT-4 on realistic tasks completed roughly 12 percent more work, about 25 percent faster, and were rated clearly higher on quality. The weakest performers gained the most. Nobody is trying to talk you out of a tool that works.

There is also a cost to the person doing the work. An MIT Media Lab study (Kosmyna and colleagues, 2025), "Your Brain on ChatGPT", measured brain activity while people wrote essays with an AI assistant, with a search engine, or with nothing. Connectivity scaled down with the amount of outside help: the no-tools group showed the strongest networks, the AI group the weakest. AI users also remembered less of what they had produced, and wrote in strikingly flat, similar ways. When the AI group later wrote with no tools, their brains still under-engaged. The researchers called this cognitive debt — a cost that does not disappear when you close the tab.

A larger survey by Michael Gerlich (2025, in the journal Societies) found more frequent AI use associated with lower critical-thinking scores, and the link ran through cognitive offloading — the habit of handing the thinking over. Younger and heavier users were most affected. Gerlich's own emphasis matters: the effect depends heavily on how the tools are used, not simply that they are used.

Now the caveats, because credibility is the point. The MIT study is a small preprint: 54 people, one task, not yet peer reviewed. The critical-thinking finding is a correlation, not proof of cause. And the idea that any of this raises dementia risk is a reasonable question, not an established finding — treat it as a hypothesis, never a fact. What we can fairly say is that independent studies point the same direction, and that is enough to act on.

Try this now. Name two tasks you regularly hand to AI. For each, decide: is it inside AI's frontier, where it is genuinely strong, or outside it, where your judgement carries the work and AI may quietly make it worse? The outside-frontier ones are where overreliance costs you most.