When the tool does the thinking
In 2025, MIT Media Lab researchers wired up 54 people with EEG and had them write essays three different ways: unassisted, with a search engine, or with ChatGPT. Across memory, attention, and executive-function networks, the ChatGPT writers showed the weakest neural connectivity of the three groups. Afterward, the researchers asked each person to quote a line from the essay they had just written. In the first session, 83.3 percent of ChatGPT users couldn't produce a single quote, correct or not. The other two groups did far better.
Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes, "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task", MIT Media Lab, June 2025. media.mit.edu
A separate 2025 study of 666 adults in the UK found a strong negative correlation, -0.68, between how often someone used AI tools and how they scored on a critical-thinking test. The author, Michael Gerlich, attributes this to cognitive offloading. When a tool does the remembering and reasoning, the person gets less practice at both. The effect was largest in 17-to-25-year-olds, the age range that includes most high school and college students.
Gerlich, "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking", Societies 15(1), 2025. doi.org/10.3390/soc15010006
Anthropic, which makes the Claude chatbot, has published data that points the same way. It sorted 574,740 real conversations between students and its own chatbot by how the student used it. Close to half, 47 percent, were what Anthropic calls Direct: the student asked for an answer or a finished piece of writing and got one back, with little exchange in between.
Anthropic, "Education Report: How university students use Claude", April 2025. anthropic.com
In all three studies, the tool did the part of the task the student was supposed to be doing.
When the tool is built to hold back
Tools set up differently have produced very different results. Harvard physics lecturer Gregory Kestin built a chatbot tutor for his course and told it to give away one step at a time and never the full solution. In a study of 194 students, the group that studied with the tutor learned about twice as much as the group in the usual in-person active-learning class, and reported roughly double the engagement.
Kestin, Miller, Klales, Milbourne and Ponti, Harvard physics tutoring study, fall 2023, reported by Harvard Gazette, September 2024. news.harvard.edu
Older research points the same way. A 2016 meta-analysis pooled 50 controlled evaluations of intelligent tutoring systems built the same way, one step revealed at a time, and found a median gain of 0.66 standard deviations over conventional instruction alone, roughly the difference between scoring at the 50th percentile and the 75th.
Kulik and Fletcher, "Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review", Review of Educational Research 86(1), 2016. journals.sagepub.com
A large field test ran in Nigeria in 2024. The World Bank gave students in nine public secondary schools six weeks of after-school English tutoring with a GPT-4 chatbot, with teachers guiding each session. The gain, 0.31 standard deviations overall, was equivalent to roughly a year and a half to two years of ordinary schooling, at a cost of about $48 per student. The largest gains were among girls.
De Simone, Tiberti, Barron Rodriguez, Manolio, Mosuro and Dikoru, "From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria", World Bank Policy Research Working Paper 11125, May 2025. documents.worldbank.org
What separates the two groups
The chatbots in the first set of studies will hand a finished essay or answer to anyone who asks. In the second set, a teacher or researcher configured the tool so the student still had to do the work. Sometimes the underlying model was the same in both. The difference was in how it was set up.
So we don't treat "AI in the classroom" as a single question. Writing a student's essay and walking a student through a problem one step at a time are different uses. The same question applies to classroom material: who did the thinking, and did anyone check the result. AI slop is already reaching schools where nobody does.
