Designing with AI: Developing Evidence-Based Models for AI Integration in Bioengineering Research Training
Students learning to design research need to engage in genuine creative and analytical work, yet AI tools can either support or undermine that process depending on how they are used. This project develops and studies a novel instructional model in which AI serves distinct pedagogical roles during bioengineering research design: AI as critic (providing feedback on student-generated ideas) versus AI as co-designer (generating ideas with and for the student). The researchers will conduct a randomized crossover experiment with ~60 students in BIOE141, a bioengineering senior capstone design course, assessing impacts on experimental design quality, diversity of ideas, learning of experimental design skills, AI tool use practices, learner agency, and motivation. This research aims to develop an evidence-based model for integrating AI into research training, with broad applicability to engineering courses within and beyond Stanford.
Research Team: Emma Lundberg, Ross Venook, Sarah Klass, Catherine Chase
TeachableAgent: Learning by Teaching!
Most AI tutors summarize, scaffold, and solve for the student. This can make instruction more accessible, but it risks making the student a passive consumer. Decades of learning science research reveals that passive exposure to clear explanations creates an illusion of competence: students feel fluent at the moment of explanation, but struggle to reconstruct the concept from first principles. More importantly, because the AI does the explaining, student misconceptions often remain hidden. This project introduces TeachableAgent, an interactive learning environment that reverses the conventional tutoring relationship by asking students to teach an AI rather than be taught by one. Grounded in the protégé effect and Feynman Technique, the system prompts students to articulate their reasoning, provides limited support when necessary, and offers feedback on misconceptions, missing reasoning steps, and areas for practice. A preliminary pilot at Stanford found that students’ explanations became more complete, precise, and coherent after using the system. These findings motivate a broader study of the approach’s effects on learning and long-term retention.
Research Team: John Mitchell, Anushree Aggarwal
Proxy-Expert Practice: An Agentic AI System for Teaching Entrepreneurial Problem Validation
Problem validation — determining whether a problem is real, important, and worth solving — is widely regarded as the most consequential skill in entrepreneurship and among the hardest to teach. This project proposes an agentic AI system that serves as a proxy-expert: an AI instantiation of a specific real domain expert’s problem representation, grounded in multi-session discovery interviews with Stanford Distinguished Careers Institute (DCI) fellows. Students practice problem validation with the proxy-expert through adversarial verification, then engage real fellows in a live transfer study. The work contributes to learning science by characterizing proxy-expert pedagogy as a new paradigm — AI as embodied problem, distinct from tutoring and population simulation — and generalizes to experiential project-based education broadly.
Research Team: Charles Eesley, Itai Ashlagi, Yikai Cao, Xilan Zhang
Human–AI Collaboration in Economics and Social Science
Researchers now use AI tools across nearly every workflow. Most researchers pick up these tools ad hoc, producing wide variance in adoption, technique, and outcomes. This two-week intensive bootcamp will teach doctoral students in economics and the social sciences to do research collaboratively with AI agents, through structured lectures, paired live coding, peer feedback, and a moonshot research project each participant produces during the bootcamp itself. The pilot will deliver the first randomized estimates of intensive AI training on doctoral research outcomes in the social sciences, as well as an open curriculum that peer institutions can reuse.
Research Team: Lukas Althoff
The Decision Body: A Generative AI Simulator of How Today’s Health Choices Shape Tomorrow’s Self
Stanford students learn about aging biology, multi-omics, AI, and ethics in separate courses. This project proposes to build The Decision Body, a generative-AI simulator that integrates these topics into a multifaceted learning tool that provides students with a personalized experience based on student data entry. The simulator returns the following: a photoreal visualization of the user’s self at five-year increments out to several decades, generated by a controllable diffusion model conditioned on validated aging biomarkers, not on naïve “make me look older” image prompts; an organ-by-organ aging trajectory for the 79 organ systems profiled in the Snyder Lab’s longitudinal multi-omics work, surfaced as interactive plots with calibrated uncertainty bands; and counterfactual scenarios — what if you slept seven hours instead of five, what if you started resistance training at 40 — showing how each decision branch shifts both the imagery and the underlying trajectories.
Research Team: Michael P Snyder, Ronjon Nag, Artem A. Trotsyuk
Translating Instructor Expertise and Learning Science Knowledge into Individualized Course Companions
Students at Stanford are increasingly turning to generative AI tools for day-to-day academic support. Without guidance, however, students risk using these tools in ways that bypass the active sense-making, deliberate practice, and productive struggles that are essential for learning. Faculty at Stanford are also concerned about students’ use of generative AI and struggle to develop teaching strategies or educational activities that responsibly support and assess students in the generative AI era. This project aims to make effective AI-supported teaching and learning scalable for faculty and students across Stanford University. Its main innovation is a novel AI-driven interactive framework that helps Stanford faculty to translate their deep domain expertise and teaching experience into learning-science-informed AI course companions for their Stanford courses. The course companions that faculty will create with the help of this new framework will address a wide range of students’ individualized learning needs while ensuring that the support is aligned with instructors’ domain expertise, instructional goals, and evidence-based teaching and learning principles.
Research Team: Candace Thille, John Mitchell, Guillermo Solano-Flores, Judith Fan, Yungsung Kim, Marcos Rojas
AI-Enabled Archival Research in the Classroom
This proposal seeks to develop and pilot an adaptable curriculum linking AI literacy, archival research, and African Studies. Built around HistoryGenie, an open-source platform for transcribing, searching, querying, and citing multilingual historical sources, the project will train students to use AI as a research partner rather than a shortcut. The unit will be integrated into Stanford humanities courses, assessed through surveys, student work, and pre-/post-measures, and shared publicly as a reusable teaching resource for evidence-based, responsible AI-supported archival research.
Research Team: Grant Parker, James T. Campbell, Rachel Jean-Baptiste, Joel Cabrita, Steven Press
The Socratic Tutor: Evaluating an AI tutor for assisting undergraduates with reading comprehension
To improve student comprehension of readings assigned outside of class, the researchers designed an AI-assisted application, the Socratic Tutor. The Socratic Tutor is a chatbot that holds a short conversation with students about their understanding of any input reading and can prompt them to re-read or re-focus on aspects of the reading that they have not yet mastered. The team has recently implemented the app in a large introductory course on cognitive science, SYMSYS1: Minds and Machines, replacing or supplementing traditional reading response assignments. In addition to being more responsive than traditional reading responses to students’ needs, the Socratic Tutor is also more resistant to student cheating using AI. The goal of the current project is to evaluate the success of the Socratic Tutor at increasing student comprehension of assigned readings and of course content more generally.
Research Team: Michael C. Frank, Huijun Mao, Bonnie Krejci
Parallel Problems by Design: Pedagogically-Structured AI Tool to Practice Transfer in Introductory Chemistry and Physics
Transfer is a cornerstone of effective learning, yet hard to achieve. Students commonly fail to transfer previously-learned knowledge to a new context. This project will examine whether AI can empower teachers to scaffold student transfer by enabling the generation of instructional materials to deliberately practice effective transfer. Instructors often find that students can follow a worked example in class but struggle when asked to solve a similar problem on their own with small, surface-level changes. Students, meanwhile, often feel that exam questions are very different from the problems they practiced beforehand in lecture or assigned problem sets. This project will examine the potential of LLMs to address this instructional gap: can LLMs assist instructors in generating parallel problems — different in surface features but similar in deep structure, for instructor targeted learning goals and student knowledge levels?
Research Team: Shima Salehi
CLARITY CVC: AI Guided Video Assessment and Feedback for Central Line Procedural Education
Ultrasound guided internal jugular (IJ) central venous cannulation (CVC) is a common and clinically important procedure and a core competency across many clinical specialties. CVC placement is directly connected to patient safety, and national guidance emphasizes both ultrasound guidance and appropriate training for clinicians who insert central venous catheters. At Stanford, central line training already includes structured preparation, simulation practice, and verification of proficiency using a checklist that defines key technical and safety standards. However, assessment remains dependent on direct human observation. The researchers propose to develop and evaluate CLARITY CVC, an AI-guided educational platform that uses procedural video to support more objective and instructional assessment of ultrasound guided IJ CVC performance. The system will record learner performances in a standardized neck-based simulation environment, analyze video using computer vision and temporal modeling, estimate competence relative to a faculty developed benchmark, and return structured feedback that connects specific video moments to procedural standards. The central hypothesis is that AI augmented, physician supervised feedback will improve learner competence, reduce safety errors, and shorten time to safe proficiency compared with standard simulation feedback alone.
Research Team: Louise Sun
