AI Helps You Write Faster, But Teaches You Less

The viral headline said AI “reprograms the brain” and causes “cognitive decline.” That is not what the study shows. The paper, from an MIT team led by Nataliya Kosmyna, is more careful and more useful. It argues that when we lean on generative AI to write, we often produce more with less mental engagement, and we remember less of what we just made. The authors call this cognitive debt, a debt that comes due when you later need to recall, explain, or build on the work without the tool. You can read the study yourself here: Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant.

AI smooths the path to a finished page, while quietly lowering the mental “work” that cements learning.

That nuance matters. It also maps to what many knowledge workers feel intuitively. AI can make you look fluent faster. It does not automatically make you competent.

The researchers ran a multi-session writing experiment. Participants drafted short essays under three conditions. One group used a conversational AI assistant. Another used a standard search engine. A third wrote unaided. Across sessions the team recorded EEG signals associated with cognitive load and engagement, interviewed participants about their process, and tested how well writers could recall or quote their own text. The essays were also evaluated for quality.

Two patterns stood out. First, AI assistance tended to reduce the neural signatures of mental effort during writing. Second, the AI-assisted group remembered less of what they had produced shortly afterward compared with the unaided group. That is a tradeoff, not a catastrophe. Lower load, faster output, weaker encoding.

There was a hopeful wrinkle. Participants who first practiced writing without AI then added the tool later used it more effectively, drawing on it to polish or to check claims rather than to generate entire passages. They retained more. The inverse was worse. Starting with AI made it harder to build the underlying skill and to internalize content.

Speed without strain feels productive today. It can leave you with less to build on tomorrow.

None of this amounts to proof of long-term “cognitive decline,” and EEG data is not a magic window into your soul. It is evidence of how a specific kind of help changes what your brain does during work, and what gets stored.

The MIT study slots into a long line of research on how tools reshape memory. A decade ago, Betsy Sparrow and colleagues showed that people remember where to find information online better than the information itself, a result nicknamed the Google effect in Science. Psychologists call this cognitive offloading, and a sweeping review by Risko and Gilbert documented how we routinely shift mental load onto notebooks, devices, and other people in Perspectives on Psychological Science.

We have lived this before. Calculators reduced the need for hand computation while changing how math is taught. GPS changed how we navigate, with measurable impacts on spatial memory and strategy use in some studies. The richer claim is philosophical. Andy Clark and David Chalmers argued that minds extend into the tools we trust, from paper to phones, in their influential essay on the extended mind published in 1998.

Generative AI intensifies all of that. Search engines help you find, calculators help you compute, GPS helps you orient. A writing model helps you think in sentences. It stands closer to the act of reasoning itself, which is why the line between assistance and substitution gets blurry fast.

There is another echo in the economics of work. Experiments with generative AI in the workplace generally find faster output and bigger gains for lower-skilled workers, along with a risk of deskilling if training is an afterthought. See Noy and Zhang’s field experiment on writing productivity from the NBER, or GitHub’s data on software developers using Copilot to complete tasks more quickly on the company’s research blog. Other studies warn of a hidden cost, from shallow understanding to insecure code when users over-trust suggestions in security experiments.

Memory is not a moral virtue. It is a strategy. New tools change our strategies, and the bill comes due in different currencies.

The right takeaway from MIT is not to ban AI. It is to use it deliberately. The problem is not help. The problem is when and how you accept it.

• Write before you prompt. Do a first pass from your own head, even if rough. Retrieval practice strengthens memory. Classic studies show testing yourself beats rereading for long-term learning in Science.
• Ask for critique, not content. Have the model highlight gaps, surface counterarguments, or check calibration. Keep authorship and reasoning with you.
• Use AI to widen, not replace, your search. Generate diverse outlines or perspectives, then go read the sources. Add links you actually vetted.
• Make a scratchpad you will revisit. Externalize what you learn in your own words. Tools can live in the loop, but the summary should come from you.
• Timebox the sugar. If you feel yourself slipping into autopilot, stop and restate the argument from memory. If you cannot, the debt is growing.

Educators and managers can help by sequencing tasks. Have people tackle core work unaided first, then layer assistance. That matches the MIT finding that prior skill buffers against debt. It also respects the reality that modern work combines recall, synthesis, and polish. AI is excellent at the last one. You should own the first two.

As for the scary headline, set it aside. Brains are plastic across a lifetime. EEG measures of engagement are not destiny. The useful question is practical. When is the extra speed worth the shallower imprint, and when is it not?

On a plane, the student who lets ChatGPT outline the deck will land with a finished file, and with less internalized understanding of what is inside. That is not doom. It is a choice. The smart habit is to decide which parts you want your tools to carry, and which parts you want your mind to keep.