'Is AI Making Us Dumber?' The Cognitive-Debt Research Every Developer Should Read

A wave of coverage this month keeps circling the same uncomfortable question: is leaning on AI quietly eroding our ability to think? Business Insider ran “Is AI making us dumber?” on July 7, and The Straits Times followed on July 21 with a guide to using AI “without losing memory, focus and thinking skills.” The concern isn’t vibes. It’s grounded in an MIT experiment that watched people’s brains while they wrote — and the group that used a chatbot came out weakest. For developers, who offload more thinking to AI than almost anyone, this is worth a hard look.

What happened

The anchor is a study out of MIT’s Media Lab, led by researcher Nataliya Kosmyna, titled “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” Researchers recruited 54 students from Boston-area universities and split them into three groups for an essay task: a “brain-only” group with no digital tools, a search-engine group (Google allowed, no chatbots), and a group that could use ChatGPT. Each person wrote under a 20-minute limit while EEG recorded their brain activity across three sessions.

The results tracked with engagement. Brain-only writers showed the strongest, most distributed neural networks. Search users landed in the middle. ChatGPT users showed the weakest brain connectivity. In a fourth session the researchers flipped the conditions — and the people who’d relied on ChatGPT the whole time struggled to write well on their own. They’d gotten used to outsourcing the work.

The team named the effect cognitive debt: like technical debt, you borrow efficiency now and pay for it later in eroded memory, creativity, and critical thinking. The catch, they note, is there’s no easy way to pay it back down. (Worth flagging: this paper is a preprint and hasn’t cleared peer review yet, and 54 people is a small sample. But it’s converging with peer-reviewed work — including a 2026 comparative study of young adults in Cognitive Processing — that points the same direction: the more you offload to AI, the less your own critical-thinking muscles do.)

Why it matters

The finding most worth sitting with isn’t about IQ. It’s that clean, confident AI output creates a false sense of understanding. Participants often felt sure of answers that were shallow or wrong, because well-structured prose reads like comprehension even when none happened. That’s the trap: the tool doesn’t just do the work, it convinces you that you could have.

What this means for developers

You are the target demographic for cognitive debt. Autocomplete finishes your functions, chatbots explain the error, agents write the boilerplate. Most of it is genuinely good. But the same loop that makes you faster can hollow out the very skills — reading unfamiliar code, holding a system in your head, reasoning through a bug — that make you a senior engineer in the first place.

The fix isn’t to quit the tools. It’s to keep a few reps for yourself:

• Draft before you prompt. Sketch the approach or the function signature in your own head first, then let AI fill it in. You keep the reasoning; it does the typing.
• Read the generated code like a reviewer, not a recipient. If you’d fail it in review from a junior, don’t merge it from a model. Understanding-before-accepting is the rep that pays down the debt.
• Keep “brain-only” blocks. Protect one focused stretch a day — a gnarly bug, an architecture decision — where you don’t reach for the assistant. Treat it like strength training for the parts of your brain the tools would happily take over.
• Track it like a habit. “Solved one problem unaided today” is exactly the kind of streak worth logging. What you measure, you protect.

“While LLMs offer immediate convenience, our findings highlight potential cognitive costs.” — Kosmyna et al., “Your Brain on ChatGPT,” MIT Media Lab

The bottom line

AI isn’t making anyone dumber on its own — unexamined reliance on it might. The developers who win the next decade won’t be the ones who prompt the most; they’ll be the ones who stayed sharp enough to know when the machine is confidently wrong. Use the tools. Just don’t let them do the one thing you’re actually paid for: thinking.