The Dean’s advice to our students is “Never allow AI to think for you.” Researchers are now measuring what generative AI such as ChatGPT does to learning. We collect the studies worth reading here, with a short summary of each, its full citation, and the paper itself to read or download.
Does AI Help or Hurt Learning?
The studies below reach different conclusions, and they point to one explanation: how a student uses AI matters. In the two experiments that examined how people used it, those who asked AI for answers lost ground once it was taken away, while those who used it for hints or to explain concepts did as well as people without AI, or better. The observational study, which follows millions of math exercises over ten years, found that students spent less time on the kind of problem AI can solve and remembered less of it when tested without AI.
These papers are preprints posted to arXiv, the open research archive run by Cornell University. A preprint may not yet have completed peer review. Each summary names the study’s design, because an observational study and a randomized experiment support different kinds of conclusions.
Randomized experiment · 211 undergraduates · First posted July 2026
Experimental Evidence on the Learning Impact of Generative AI
Zara Contractor and Germán Reyes, Middlebury College
In proctored, in-person sessions, undergraduates learned about an unfamiliar topic (blockchain, carbon capture or CRISPR) and wrote an analytical essay, with or without access to a generative AI tool. All of them then took tests without AI, once immediately and again one week later.
Students with AI access scored higher on the immediate knowledge test, by 0.27 standard deviations, and most of that gain remained a week later. Their essays changed little while they had AI, but the essays they wrote a week later without AI were better in style and relevance to the question. Those later gains were larger among students who had used AI to explain concepts than among students who had used it to write text for them.
Download the Paper (PDF) View on arXiv
Citation: Contractor, Z., & Reyes, G. (2026). Experimental evidence on the learning impact of generative AI (arXiv:2607.08849v1). arXiv. https://doi.org/10.48550/arXiv.2607.08849
License: Creative Commons Attribution 4.0 (CC BY 4.0). The copy offered here is version 1 from arXiv, with its file size reduced by Godhunting Resources for faster download; the text and figures are unchanged.
Observational study · 3.2 million math learning interactions · First posted May 2026
Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build
Sina Rismanchian (University of California, Irvine), Hasan Uzun, Jeffrey Matayoshi, Eric Cosyn and Eyad Kurd-Misto (McGraw Hill)
The authors examined ten years of records from ALEKS, an online mathematics learning program, comparing word problems that AI can solve easily with graph-based problems it handles poorly. After ChatGPT’s release, the time students spent on the AI-susceptible problems fell by a cumulative 26.9 percent among college students and 31.3 percent among high school students.
On proctored assessments, where AI could not be used, the odds of a correct answer on the AI-susceptible kind of problem declined by a cumulative 25 percent. The authors read this as students finishing problems faster with AI without building the knowledge the problems were meant to teach. Because the study observes a change after ChatGPT’s release rather than assigning students to use AI, it shows a pattern consistent with that explanation, not a controlled test of it.
Download the Paper (PDF) View on arXiv
Citation: Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster completion, less learning: Generative AI reduced study time on math problems and the knowledge they build (arXiv:2605.21629v3). arXiv. https://doi.org/10.48550/arXiv.2605.21629
License: Creative Commons Attribution 4.0 (CC BY 4.0). The copy offered here is version 3 from arXiv (July 13, 2026), unmodified.
Three randomized experiments · 1,222 participants · First posted April 2026
AI Assistance Reduces Persistence and Hurts Independent Performance
Grace Liu (Carnegie Mellon University), Brian Christian and Tsvetomira Dumbalska (University of Oxford), Michiel A. Bakker (MIT) and Rachit Dubey (UCLA)
Participants worked fraction problems or reading-comprehension questions, some with an AI assistant and some without. After about ten minutes of help, the AI was taken away. Those who had used it did better while they had it, but afterward they solved fewer problems on their own than participants who never had it, and they gave up on problems more often.
The harm was concentrated among participants who asked the AI for direct answers. Participants who used it for hints performed about as well as those who worked without AI.
Download the Paper (PDF) View on arXiv The Authors’ Illustrated Summary
Citation: Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance (arXiv:2604.04721v4). arXiv. https://doi.org/10.48550/arXiv.2604.04721
Offered here with the permission of the corresponding author, Grace Liu. The copy is version 4 from arXiv (August 5, 2026), unmodified.
