Does AI Help or Hurt Learning? *
Open accessAbstract
AI is transforming how students learn, raising concerns about whether it expands educational opportunities or widens existing gaps. We examine this question in a preregistered lab experiment (N=572) in which students study a novel topic under one of three conditions: browsing only (control), AI-assisted, or AI-guided, and then complete an exam without AI access. We find no overall effect of AI access on learning outcomes. However, this average zero effect masks substantial heterogeneity. High-GPA women appear to benefit the most from AI-guided access, while the effects on men and low-GPA students are weaker and in some cases negative. We also find that students with AI access attempt fewer practice questions during the study phase. This suggests that studying with AI crowds out other learning activities, but does not lead to an overall change in exam performance. Finally, exploratory analyses of prompt data provide suggestive evidence on why some students benefit more than others. More delegative AI use (measured by copy-pasting practice questions into the chatbot) is associated with attempting more questions but performing worse on the final exam. High-GPA women rely on this strategy the least and perform the best. Overall, AI appears to crowd out some traditional study effort without reducing learning on average, but because its benefits are concentrated among already advantaged students, it may reinforce existing educational inequalities.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In Franco et al. (2026), neither unrestricted nor guided ChatGPT access improved unassisted exam performance compared to a browsing-only control (0.04 SD for unrestricted, -0.12 SD for guided).
“neither AI arm differs from the control (0.04 SD for unrestricted access, -0.12 SD for guided access)”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al. (2026)
The Introduction and §4.1 report that students took an exam without external aids and that neither AI treatment had a statistically distinguishable effect on exam performance relative to the Google Search control. The passages available from our copy do not establish the cited point estimates of 0.04 SD and −0.12 SD; the visible table reports test-score estimates in a different form. (LLM, conf=0.81)
Experimental — accuracy not yet measuredWritten guidance does not raise exam scores (Franco et al., 2026).
“The exceptions are a scaffolded tutor whose knowledge gains are indistinguishable from ChatGPT's (Bassner et al., 2026) and written guidance that reduces question pasting without raising exam scores (Franco et al., 2026).”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al., 2026
The divergence between assisted output and unaided learning found for automation users recurs across settings.
“The divergence between assisted output and unaided learning that we find for automation users recurs across settings.”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al., 2026
Written guidance reduces question pasting (Franco et al., 2026).
“The exceptions are a scaffolded tutor whose knowledge gains are indistinguishable from ChatGPT's (Bassner et al., 2026) and written guidance that reduces question pasting without raising exam scores (Franco et al., 2026).”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al., 2026
Four studies compare learners by their observed use of the same tool.
“We also identified four studies that compare learners by their observed use of the same tool, as we do (Liu et al., 2026; Shen and Tamkin, 2026; Franco et al., 2026; Strömberg et al., 2026).”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al., 2026
Heavy copy-pasters score below the control mean (Franco et al., 2026).
“heavy copy-pasters score below the control mean (Franco et al., 2026)”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al., 2026
Comparisons by observed use of the same tool point the same way as the paper's comparisons.
“Comparisons by observed use of the same tool point the same way as ours.”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al., 2026
In Franco et al. (2026), students with AI access attempted fewer practice questions.
“students with AI attempt fewer practice questions.”
From Experimental Evidence on the Learning Impact of Generative AI (2026) · cites this paper as Franco et al. (2026)
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
No citation lists for this paper
No citation lists for this paper yet — checked again daily.
Research works printed but not linked: 5
- Dai, Xusheng, Zhaochun Wen, Jianxiao Jiang, Huiqin Liu, and Yu Zhang , 'How students arXiv preprint arXiv:2505.08672 , 2025. · arXiv
- Third Way Report 20013. , David Figlio, Krzysztof Karbownik, Jeffrey Roth, and Melanie Wasserman , 'Males at the tails: How socioeconomic status shapes the gender gap,' The Economic Journal , 2023, 133 (656), 3136-3152. Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman , 'Generative AI without guardrails can harm learning: Evidence from high school mathematics,' Proceedings of the National Academy of Sciences , 2025, 122 (26), e2422633122. Bengio, Yoshua, Stephen Clare, Carina Prunkl, Maksym Andriushchenko, Ben Bucknall, Malcolm Murray, Rishi Bommasani, Stephen Casper, Tom Davidson, Raymond Douglas et al. , 'International ai safety report 2026,' arXiv preprint arXiv:2602.21012 , 2026. Carvajal, Daniel, Catalina Franco, and Siri Isaksson , 'Will Artificial Intelligence Get in the Way of Achieving Gender Equality?,' NHH Dept. of Economics Discussion Paper , 2024, (03). use AI feedback matters: Experimental evidence on physics achievement and autonomy,' · arXiv
- DeKeyser, Robert , 'Implicit and explicit learning,' The handbook of second language acquisition , 2003, pp. 312-348.
- Han, Xiaoli, Hongchao Peng, and Mingzhuo Liu , 'The impact of GenAI on learning outcomes: A systematic review and meta-analysis of experimental studies,' Educational Research Review , 2025, p. 100714.
- Humlum, Anders and Emilie Vestergaard , 'The Adoption of ChatGPT,' University of Chicago, Becker Friedman Institute for Economics Working Paper , 2024, (2024-50). Kestin, Greg, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti , 'AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting,' Scientific Reports , 2025, 15 (1), 17458. Koutny, Ilona , ' A typological description of Esperanto as a natural language,' Język. Komunikacja. Informacja , 2015, 10 , 43-62. Lehmann, Matthias, Philipp B Cornelius, and Fabian J Sting , 'AI Meets the Classroom: When Does ChatGPT Harm Learning?,' arXiv preprint arXiv:2409.09047 , 2024. Mishra, Punya, Danah Henriksen, Lauren J Woo, and Nicole Oster , 'Control vs. agency: Exploring the history of AI in education,' TechTrends , 2025, 69 (2), 247-253. OECD , 'The ABC of Gender Equality in Education. Aptitude, Behavior, Confidence,' 2015. Pallant, Jessica L, Janneke Blijlevens, Alexander Campbell, and Ryan Jopp , 'Mastering knowledge: The impact of generative AI on student learning outcomes,' Studies in Higher Education , 2025, pp. 1-22. Pardos, Zachary A and Shreya Bhandari , 'ChatGPT-generated help produces learning gains equivalent to human tutor-authored help on mathematics skills,' Plos one , 2024, 19 (5), e0304013. Tan, Steffi and Vaikunthan Rajaratnam , 'Critique of Generative AI Can Harm Learning Study Design,' Available at SSRN 4898213 , 2024. Voyer, Daniel and Susan D Voyer , 'Gender differences in scholastic achievement: a metaanalysis.,' Psychological bulletin , 2014, 140 (4), 1174. Wecks, Janik Ole, Johannes Voshaar, Benedikt Jost Plate, and Jochen Zimmermann , · arXiv
- Authors
- Franco, C., Irmert, N., & Isaksson, S.
- Affiliations
- Norwegian School of Economics (NO)
- Lund University (SE)
- ESC Dijon Bourgogne, École supérieure de commerce de Dijon-Bourgogne (FR)
- Corresponding author
- Isaksson, Siri
- Venue
- SSRN Electronic Journal
- Publisher
- Elsevier BV
- Published
- 2026
- ISSN
- 1556-5068 (ISSN-L)
- OpenAlex
- W7151072486
- Cited by
- 0
- References
- 0
- Type
- posted content
- Language
- en
- Open access
- green
- doi.org
- Locations
- SSRN Electronic Journal (acceptedVersion)
- Lund University Publications (Lund University) (submittedVersion, other-oa)
- Funders
- Jan Wallanders och Tom Hedelius Stiftelse samt Tore Browaldhs Stiftelse
- Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung
- Topics
- Physical Sciences › Computer Science › Information Systems › Artificial Intelligence in Education
- Physical Sciences › Computer Science › Artificial Intelligence › Intelligent Tutoring Systems and Adaptive Learning
- Physical Sciences › Computer Science › Computer Science Applications › Online Learning and Analytics
- Keywords
- AI-assisted learning, educational inequality, student performance, gender differences, GPA, chatbots, educational technology
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
