Stanford Graduate School of BusinessIFDM TPF Session 3: AI in Financial Education
Stanford Graduate School of Business
Introducing The Session 0:07
The session opens with introductions at Stanford, where the host notes that AI in education is a major topic of excitement, especially given the Silicon Valley setting. Angela Amarillas, director of Stanford's Mind Over Money financial wellness program, is introduced as moderator. She frames the panel as a practical, evidence-based discussion on using AI in financial education, introducing panelists Hakan Ozas, Stanford's experimental research director who led a randomized controlled trial on AI in teaching, along with co-investigators Daniel Chi from University of Nevada Las Vegas and Jodi Leovage from Cal State Northridge, who both taught treatment and control sections in the study.
Testing AI In Classrooms 4:30
Hakan Ozas explains the project's goal: improving financial education by addressing the problem that personal finance courses are too uniform, when students arrive with very different financial backgrounds. Generative AI offers a way to personalize learning, but risks encouraging over-reliance and weakening critical thinking, since productive struggle, meaning making mistakes while learning, matters for retention. Rather than giving students open access to tools like ChatGPT or replacing instructors, the study built ten structured educational chatbots, one per personal finance topic, tested at Cal State Northridge and University of Nevada Las Vegas as a complement to normal teaching, comparing treatment and control sections taught by the same instructors.
Designing The SPARK Chatbots 9:02
The chatbots followed design principles Jodi named SPARK, standing for structured, personalized, adaptive, reflective, and the goal of avoiding direct answers. Each bot tied to a clear learning goal, such as comparing two insurance plans or deciding whether a lower deductible premium is worth it, used personalization by asking students about their own finances, such as how much they could realistically save next month, and adapted when students answered incorrectly by nudging them toward better questions rather than correcting them outright. The reflective piece applied a Socratic method, with bots instructed never to give direct answers so students had to reason through decisions themselves.
Measured Gains In Knowledge 11:31
Results showed financial knowledge test scores rose about six percentage points in control sections, while AI sections roughly doubled that gain, an effect size comparable to the typical 20 percent of a standard deviation improvement seen between receiving financial education and not receiving it at all. Students also showed better judgment about when to trust AI versus their own reasoning, and no increase in over-reliance was detected, meeting the study's central design goal. Results on applying knowledge were mixed, with improvements seen in index funds and mortgage decisions but not insurance, and over the three to four month semester no broad behavioral change appeared, though there was suggestive evidence of reduced costly borrowing, with long-term data collection still ongoing.
Demonstrating A Live Chatbot 15:01
Jodi demonstrates one of the chatbots, a car-buying tool designed to walk students through a realistic financing scenario. It begins with an introductory explanation that it is not a sales tool, won't tell the student what to buy, and won't make decisions for them, but instead helps them see trade-offs clearly, covering monthly payments, insurance, gas, maintenance, and how changes in APR or loan length can add thousands of dollars over time. In the live interaction, Jodi role-plays a student unsure of what car to buy, and the bot gently guides her toward naming a smaller car, a Honda, with a budget under fifteen thousand dollars, before continuing the structured walkthrough.
Chatbots push students to think harder 18:30
One chatbot example walks a student through buying a car, asking what matters most, their income, and what city they will drive in, then surfacing costs they had not considered like parking, car permits, and insurance. The point was never to hand over an answer but to probe the student's thinking and connect general advice to their real situation.
UNLV's financial literacy program at scale 20:01
Daniel Chief explains that UNLV now enrolls over 1,574 students per semester in financial literacy, roughly 3,000 a year, taught in small capped classes of 60 by a mix of full time staff and adjuncts drawn from bankers, the Nevada state treasurer, and CFA society members. With such a large and diverse student body, ages 16 to 82, instructors cannot have one on one conversations with everyone, which is where adaptive AI tools help deliver personalized guidance at scale.
Structured AI drives deeper engagement 24:30
Treated students engaged more, not less, because the AI interactions were deliberately structured to force thinking rather than just answering questions, producing what the researchers call productive struggle. Knowledge gains were clear, but behavioral competence was mixed, since acting correctly on a mortgage or insurance decision is far more complex than knowledge alone, especially compared to simpler choices like comparing index fund expense ratios.
Classroom evidence and instructor feedback 28:30
Comparing sections with and without chatbot access, one instructor found students using the tool produced far sharper, more coherent smart goals tied to their actual budgets. Teaching evaluations also showed students appreciated chatbots adjusting to their personal knowledge level, which reduced intimidation and made them more willing to approach instructors afterward.
Privacy and personalization advantages 31:00
Because money is culturally a taboo topic alongside death, sex, and taxes, students often hesitate to discuss finances with peers or instructors, but a judgment free chatbot lets them engage more openly, with guidance against sharing things like social security numbers. On personalization, a housing chatbot was built with different scenarios for different locations and costs, recognizing that generic advice like the national average home price is useless to a student in high cost Los Angeles versus one in Iowa.
Teaching With a Structured AI Tool 35:01
The panel describes AI as the new landscape students already live in, so courses should teach them to use it well, check sources, and build AI literacy rather than ignore it. At UNLV, where many students work, are first-generation, and vary widely in age and background, a structured chatbot lets the course personalize learning in ways a single classroom lecture cannot. Early data show no gender or demographic gap in knowledge gains, though students who already used AI heavily before the course benefited most. Rollout problems included memory glitches and students initially seeing it as extra work, solved by framing the tool as genuinely useful and introducing it gradually rather than all at once.
Assessment, Design, and Sharing the Bots 45:01
Audience questions pushed back on whether removing graded assessment removes the pressure that drives real learning, with one panelist answering that roughly eighty percent effort should go to learning and twenty percent to reinforcing assessment. The bot interactions are sequenced week by week per topic, using a badge system requiring correct answers before advancing, with interaction data still being analyzed for quality, not just completion. Researchers acknowledged instructor variation hasn't yet been isolated statistically. They confirmed plans to make the chatbots publicly available once the study concludes at the end of the academic year.
Grades as an Incentive to Engage 52:30
One presenter explains that tying a portion of the grade, about 20 percent, to chatbot interactions likely explains why usage stayed steady rather than dropping off after frustrating moments. That incentive level is something instructors can adjust based on their own preferences, and it seems to be a key reason students kept using the tool instead of abandoning it.
Designing Chatbots That Do Not Give Answers 54:30
The team stresses that their chatbots never simply hand over answers, unlike turning to ChatGPT for a quick solution. Instead the bots walk students through reflecting on themselves, since people enjoy learning about themselves, and even a modest 20 to 30 percent improvement over zero learning from copied answers is considered a meaningful gain.
Winning Over AI-Resistant Students 55:02
Asked about students who resist AI for pride, privacy, or principle, the presenter says the tool is introduced as something experts built specifically for the student, not a generic send-off to ChatGPT or Claude, complete with guardrails like never uploading bank statements or social security numbers. The framing is that AI is here to stay, much like the internet, so the goal is teaching responsible use rather than forcing adoption.
Gamification and Critical Thinking 57:01
The team stumbled into gamification partway through and added badges and levels after students responded positively, calling it a small dopamine hit. Critical thinking was built in by training the chatbot to probe students with questions rather than give answers, pushing them to reason toward a conclusion instead of just receiving one.
Same Material, Added Tool 59:01
Asked whether AI let them cover more material, the presenters clarify that treatment and control sections cover identical content; the only difference is that treatment students get the extra chatbot modules layered onto the same coursework, giving them another way to engage more deeply rather than more material overall.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
