Stanford reviewed over 50 studies of AI in school. Students fared better when they did the thinking first.
Joe, founder of FinBizify
·
Stanford’s SCALE Initiative reviewed more than 50 studies of generative AI in school, and its November 2026 update finds that AI appears better for learning when students use it for learning support than when they use it to produce their work.[1] Most of the new studies involve college students, and a research agenda released two days later says “very few examine U.S. K-12 settings at scale.”[2]
What the studies found
Students who lean on AI to do the work can end up remembering less. In one randomized trial of college students, those who studied with ChatGPT did worse on a surprise test six weeks later than those who studied without it.[3] A large analysis of online math practice found a similar pattern: after ChatGPT arrived, college students did worse on proctored questions of the kind it answers easily.[4]
Enjoying AI is also different from learning from it. Students rated AI feedback as clearer and more motivating than feedback from teachers, in studies where the learning came out about the same.[1]
How the tool is built matters. In a large middle-school math study, students working with an AI tutor tried fewer problems but got more of them right. The most promising results came when the tutor was paired with a requirement to show mastery before moving on, though the evidence was suggestive, not conclusive.[5]
Tools for teachers carry risk too. In a school study in Turkey, teachers given an AI tool saw their students’ motivation drop. The review notes that teachers mostly used what the AI produced with little editing, and that the study does not prove that caused the drop.[6]
One expert’s personal read points the same way. In a newsletter of his own, David Deming, an economist and the Danoff Dean of Harvard College,[7] says he argued in January that AI removes the struggle learning needs. After reading more than 75 experiments, he reports that most showed benefits, with the clearest gains when students think first and AI guides them, and the weakest when students use it unguided.[8] A Middlebury College experiment shows the split: students who used AI as a tutor kept a gain a week later, while the gain faded for students who had AI write for them.[9]
What the evidence leaves out
Much of this research is early. Stanford notes that its collection is mostly preprints that have not been peer reviewed, and only five of the new student studies took place in U.S. K-12 schools.[1] None of these sources looks at money education.[1][2][8] The agenda lists the big questions as open: whether students keep what they learn once AI is taken away, and whether AI supports or replaces human help. Its standard is that AI’s place in education “should be earned, not assumed.”[2]
What this means for a teen learning about money
Teenagers learning to read a paycheck or a company’s margins fall outside all of this research. The cautious reading it supports is to keep AI from doing the student’s work and to judge any tool by what a student can do later without it, the same outcomes test an earlier note traced to the U.S. Department of Education. Retention is already a weak point in financial education, as another note covered.
FinBizify has no AI feature a student types into today. Its lessons draw on public filings and federal data, each figure carries a source and a date, and a person reviews each lesson before it publishes. Any AI coaching it adds later is planned to source only the reviewed lessons and to be judged by what students can do afterward. The intention is to try what is the consensus best approach, not release a shiny new feature if not conducive or imperative to central learning.
Sources
Seth Walker, Lily Fesler, JP Martinez Claeys, Chris Agnew and Susanna Loeb, “The Evidence Base on AI in K-12: A November 2026 Review,” SCALE Initiative, Stanford Graduate School of Education, November 10, 2026. https://scale.stanford.edu/ai/evidence-base-review/evidence-base-ai-k12-november-review
Carly D. Robinson, Chris Agnew and Susanna Loeb, “A Forward-Looking Research Agenda for AI and Learning: Building the Evidence Base for Genuine K-12 Transformation,” SCALE Initiative, Stanford Graduate School of Education, November 12, 2026. https://scale.stanford.edu/ai/research-agenda/forward-looking-research-agenda-for-ai-and-learning
Andre Barcaui, “ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention,” Social Sciences & Humanities Open, 2025. https://doi.org/10.1016/j.ssaho.2025.102287
Sina Rismanchian, Hasan Uzun, Jeffrey Matayoshi, Eric Cosyn and Eyad Kurd-Misto, “Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build,” arXiv preprint, 2026. https://doi.org/10.48550/arXiv.2605.21629
Philip Oreopoulos, Michael Liut, Alp Sungu and Nina Low, “Making AI Tutoring Productive: Evidence from a Mastery-Based Math Practice Experiment,” NBER Working Paper 35621, 2026. https://doi.org/10.3386/w35621
Alp Sungu, Benjamin Lira and Angela Lee Duckworth, “Generative AI Can Harm Teaching,” SSRN working paper, July 2026. https://scale.stanford.edu/ai/repository/generative-ai-can-harm-teaching
“David Deming named Harvard College dean,” Harvard Gazette, May 13, 2025. https://news.harvard.edu/gazette/story/2025/05/david-deming-named-harvard-college-dean/
David Deming, “The education of an AI skeptic,” Forked Lightning, November 16, 2026. https://forklightning.substack.com/p/the-education-of-an-ai-skeptic
Zara Contractor and Germán Reyes, “Experimental Evidence on the Learning Impact of Generative AI,” arXiv preprint, 2026. https://doi.org/10.48550/arXiv.2607.08849
FinBizify is building money education lessons that teenagers actually want to finish, taught through real companies and real dollar amounts.