People who wrote essays with a chatbot showed up to 55% less flow of information between brain regions than people who wrote alone. That flow is called brain network activity. The finding comes from an MIT Media Lab study that measured brain activity and memory while people wrote short essays.
The study followed 54 participants aged 18 to 39 from five Boston-area universities, most of them students. Each was assigned at random to one of three setups of 18 people. One group wrote with a ChatGPT account and no other websites or chatbots. A second, the Google group, searched the web with AI answers filtered out. The third had no tools at all, only their own minds. Each person kept the same setup for three sessions over four months, with a 20-minute limit on every essay.
A headset with sensors on the scalp recorded brain activity in every session. Writers working alone had the strongest and widest-ranging brain network activity. The Google group showed 34% to 48% less of that activity, and the chatbot group sat lowest. The more outside help writers had, the less information their brain regions passed to each other.
The chatbot group also fell behind on memory. After each session, researchers asked writers to quote a sentence from the essay they had finished minutes earlier, without looking. In the first session, none of the chatbot users gave a correct quote, while 15 people in the Google group and 16 writing alone did. The difference shrank once writers knew the question was coming. Even so, in the third session 6 chatbot users still could not give a correct quote, against 2 people in each of the other groups.
A fourth session took the chatbot away without warning. Only 18 of the 54 participants came back, and nine of them had written with the chatbot for three sessions. Writing with no tools, those nine showed weaker brain network activity, which the authors describe as under-engagement.
The researchers posted the paper online before outside experts reviewed it, and lead researcher Nataliya Kosmyna called its conclusions preliminary, to be treated with caution. All of its results come from one task, essay writing. The study does not say what happens when people use ChatGPT, Gemini or Claude for a long list of tasks.
