Stanford’s AI tutors guide students through problems instead of supplying answers

Stanford's STEPS chemistry tutor asks students questions to help them work through a problem rather than supplying the answer. Short for Science Tutor for Effective Problem-Solving, it is designed to preserve the reasoning practice that can disappear when a student delegates an exercise to a general chatbot. The instructional goal is a learner who can tackle the next problem, not simply a completed assignment.[1]

Stanford's Center for Teaching and Learning has been collecting examples of faculty-built AI tools through a series of community presentations. In an account published on 28 October, the university described four classroom implementations and the trade-offs their teachers encountered. The projects are notably different from the idea of deploying one general-purpose chatbot for every course. Each begins with a specific instructional bottleneck and asks what AI could do without taking the teacher or the learner out of the process.[1]

STEPS was developed by education professor Shima Salehi with senior chemistry lecturer Jennifer Schwartz Poehlmann, drawing on earlier work with physicist and education researcher Carl Wieman. Paper-based problem-solving guides helped structure students' reasoning, but it was difficult for teaching staff to provide individual feedback at scale. The AI version attempts to sustain that guidance without instantly solving the exercise. Salehi reported that students initially disliked being denied direct answers but later came to value the attention to the process. That observation is instructor-reported classroom feedback, not a randomised estimate of improved exam scores.[1]

A question, rather than a translation

The same principle appears in a course far from chemistry. Ariel Stilerman, who teaches East Asian languages and cultures, developed BungoBot with students for reading classical Japanese. Bungo is a historical literary language without living native speakers; acquiring enough grammatical skill to read manuscripts directly normally takes sustained study. A generic language model that immediately translates a passage may let a student finish an assignment without learning how to recognise the grammatical choices inside it.[1]

BungoBot instead poses targeted questions about the text. Stilerman said students began proposing changes to the tool as they used it, and the class eventually worked on calligraphic manuscripts containing poems by the historical tea master Ōtagaki Rengetsu. The team later built a reverse-lookup dictionary and a conversational simulation of an eleventh-century court lady, each serving a different learning activity. This is a documented pedagogical experiment, but the account contains no matched control group demonstrating how much of the students' progress was caused by the software rather than intensive teaching and unusual access to primary manuscripts.[1]

Other Stanford instructors used AI to solve different coordination problems. Amit Kaushal, a clinical associate professor, built a system that collects audience questions during a lecture, groups them into themes and helps identify which questions deserve attention first. It is not intended to be the student's conversational partner throughout the session. Kaushal specifically wanted to preserve direct interaction between teacher and audience instead of allowing a class-wide chat interface to become the focus of the room. Anonymous question submission can help students who do not wish to interrupt, though the summary still requires a lecturer's judgement about what should be answered.[1]

The teacher remains the final editor

In biology, senior lecturer Waheeda Khalfan and colleagues faced a different problem: students entered foundational courses with highly varied prior knowledge, and the course sequence allowed considerable flexibility. They used an AI-assisted pipeline to examine years of curriculum material and map which concepts students needed before each class. Human experts then reviewed and refined the relationships and developed early self-diagnostic assessments and preparation guides. A web-based Foundations Navigator helps students understand these dependencies and plan an appropriate route through the course series.[1]

Khalfan reported that work previously taking months could be completed in weeks, while insisting that subject experts remain responsible for the final material. This is an estimate of production workflow time from the team, not a measured reduction in student dropout rates or evidence that incorrect prerequisite maps have become impossible. It illustrates where generative AI may add value as a drafting and analysis tool while human curation still determines whether the advice is educationally sound.[1]

Across the projects the most important design choice is not which model powers the system but where authority and effort reside. STEPS keeps mathematical work with the learner. BungoBot encourages interpretation instead of outsourcing translation. The lecture moderator gives the instructor a better view of student questions. The biology navigator accelerates a curriculum-mapping job that teachers still check. These are four different answers to four different teaching problems, not examples of a single universal AI education product.[1]

Stanford says 23 instructors have presented their projects across two academic years. The university account documents approaches and classroom experience, rather than an aggregate controlled test of learning gains. Comparing later, unaided performance against suitable teaching alternatives would help establish which designs build transferable skills and which chiefly make the immediate coursework easier.[1]

Report an error or suggest a correction