- Was the Bastani study on generative AI and learning corrected or retracted?
- What is the DOI of the correction to Bastani et al. (2025)?
The Bastani study of generative AI and school mathematics has not been retracted. PNAS issued one correction notice, published online on 20 August 2025, and as the evidence base here records it the notice corrected a printed affiliation and nothing else. The finding stands as published.
The answer, in one line
No. PNAS published one correction notice for it, online on 20 August 2025, and no retraction. The correction is catalogued in Crossref as an update to the original article, and its record carries the type correction.
Definition#
Correction notice: a journal's published statement that an error in an article has been fixed, with the article left in the record. A retraction withdraws the article. The Bastani paper has the first and not the second.
What the record shows about the notice#
The paper is Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman, Generative AI without guardrails can harm learning: Evidence from high school mathematics, Proceedings of the National Academy of Sciences, volume 122, issue 26, article e2422633122, published 25 June 2025 after acceptance on 5 May 2025. Crossref lists six authors in that order.
The correction is titled "Correction for Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics". Crossref gives its DOI as 10.1073/pnas.2518204122, an online date of 20 August 2025 and a print date of 26 August 2025, in volume 122, issue 34. Its record is typed as a correction and linked as an update to the original paper. There is no retraction record attached to the original.
What the notice changed, and how far to trust that description#
The evidence base on this site records the notice as correcting the affiliation printed for Osbert Bastani, a production error, with no change to the author list, the method, any figure or any finding. That description comes from reading the notice at source on 12 September 2026. On 26 October 2026 the PNAS page sat behind a bot check and PubMed returned a verification screen, so the text could not be read a second time. The title, DOI and dates above were re-confirmed from Crossref that day; the account of what was corrected rests on the earlier reading. Anyone who needs the wording should open the notice at the DOI.
Why the finding is unaffected on the published record#
The experiment ran in a high school with nearly 1,000 students and three conditions: an unrestricted ChatGPT-style tutor, a version with prompts built to protect learning, and no AI. Access improved performance on practice problems. When access was removed, students who had used the unrestricted tool did worse than students who had never had it, and the safeguarded tutor largely prevented that drop. The authors attribute the harm to students using the tool as a crutch while practising. None of that is touched by a corrected affiliation.
Reading a correction as a warning sign is a reasonable instinct, and the check takes two minutes: look at the type of notice, then at what it says it changed. A retraction or an expression of concern would change how the paper should be used. A correction to a byline does not. The paper remains a single field experiment in one school, on mathematics, over a bounded period, which is the limit on what it can carry, and the same limit applies with or without the notice.
Where the finding is used on this site#
The paper carries a peer-reviewed grade in the evidence base and underpins the pages on how humans learn with AI and whether using AI stops you learning. The wider question it raises, depending on a tool past the point where you could catch its errors, is set out under over-reliance.
Key sources
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122
- Correction for Bastani et al. (2025). PNAS, 122(34). DOI 10.1073/pnas.2518204122, recorded in the graded entry for the paper
Related SuperSkills research#
How humans learn with AI, does using AI stop you learning, what is over-reliance on AI and what is deskilling.
About this research#
Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page reports a journal record and makes no claim of its own. Last reviewed: 26 October 2026.
Essay · SS-2026-437 · How I use AI here · 1 peer-reviewed study
Hirji, R. (2026). Was the Bastani AI learning study corrected?. The SuperSkills evidence base, SS-2026-437. https://thesuperskills.com/research/was-the-bastani-ai-learning-study-corrected. Last reviewed 26 October 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work