An AI tutor can give a student the right answer in seconds. A group of Stanford teachers is investigating whether the more valuable assistant is one that refuses to do so, at least until the student has tried to understand the problem. Their approach uses questions, prompts and carefully designed exercises to keep learners working rather than simply copying a result. It offers a different idea of what useful AI in education could look like.
The university’s October 2026 account of its teaching experiments describes several projects, including a chemistry tool called STEPS and a classical Japanese assistant named BungoBot. In both cases, the AI is intended to guide thought rather than replace it. These are accounts of developing classroom tools and student experiences, not proof from a large controlled trial that every student learns more using them.
Why an AI tutor should sometimes slow you down
Consider a chemistry problem involving several stages. A student who asks a chatbot for the answer might receive a polished sequence of calculations and believe they understand it. That feeling is not the same as being able to solve a different problem tomorrow. If the assistant has made all the decisions, there may be little opportunity for the student to practise noticing which information matters or choosing a method.
Stanford’s STEPS project was designed with a different rhythm. Developed through work involving education researcher Shima Salehi and chemistry lecturer Jennifer Schwartz Poehlmann, it asks questions intended to move students towards their own solution. The aim is to preserve what educators call productive struggle: the effort of figuring something out rather than having the difficult part removed.
Struggle can be useful, but only when there is a reasonable chance of progress. Leaving a confused student alone with a task they cannot interpret may create frustration rather than learning. An effective tutor needs to judge when to provide a hint, when to ask for an explanation and when to make a concept clearer. Those are teaching decisions, not merely differences in the number of words a system produces.
Stanford reports that some students initially disliked STEPS because it would not give them an immediate answer. Later feedback was more positive about learning a method for solving problems. That is an interesting change in perception, but it should not be promoted into a measured improvement in grades or long-term understanding without appropriate comparative evidence.
The difference between a hint and a completed assignment
There is a practical distinction between ‘What is the answer?’ and ‘Why did I choose the wrong first step?’ An assistant could respond to the second question by examining the student’s reasoning. It might identify a mistaken assumption, suggest an example or ask the learner to explain the result in their own words. The student remains responsible for producing the solution.
That may feel less efficient in the moment. It can involve more typing, more errors and a longer conversation. For someone who only needs a one-off factual answer, the extra work may be unnecessary. For someone who wants a transferable skill, however, time spent thinking is not always wasted time. The educational purpose should determine how much assistance is appropriate.
The contrast matters because schools and universities have spent years debating what happens when students rely on AI to complete ordinary work. LiveAIWire recently examined research on how ready access to AI assistance can reduce persistence on independent tasks. That study raised a concern about behaviour after help is removed. Stanford’s designs are a potential response, not a demonstrated solution to every problem identified elsewhere.
Teachers can also design assignments so that the explanation matters. An answer that looks correct may have been guessed or generated. Asking a student to explain why a method works, compare two approaches or identify an error creates more evidence about understanding. An AI tutor that supports those activities could be useful, provided it does not write the explanation for the learner.
A classical Japanese lesson gives the idea another test
Stanford’s BungoBot applies a questioning approach to classical Japanese. Instead of immediately translating a difficult sentence, it prompts students to examine the language and reason towards a reading. That matters because the course asks students to acquire specialist interpretative skills, not merely receive an English version of a text.
The university describes positive student engagement and rapid iteration as the instructor and class improved the tool. Students’ comments helped adjust how it responded, and the assistant became part of the learning process. Such feedback is valuable for refining a classroom experience, but it does not establish how the same system would perform with different teachers, institutions or groups of learners.
The subject also illustrates that AI tutoring is not one product category with a single ideal script. A language student may need a prompt about grammar or historical context. A science student may need a question about assumptions in an equation. A pupil learning to write may benefit from suggestions about structure while retaining authorship of the sentences. The common principle is assistance designed around the skill being learned.
What independent education research adds
A 2026 expert consensus project at the University of Minnesota examines ways generative AI may support or threaten learning across education. It treats benefits and risks as dependent on how systems are used, rather than assuming that all AI help is either constructive or damaging. That is a useful frame for judging Stanford’s work without turning promising demonstrations into a blanket recommendation.
An educator needs to ask what happens after the assistant disappears. Can the student retrieve the idea without prompting? Can they explain it, adapt it and notice when it does not apply? These are more demanding tests than whether a learner enjoyed using a tool or produced a correct answer while the tool was available.
There are also practical constraints. A tutor that asks questions must ask relevant ones. If it confidently pursues a false assumption, a learner may spend time following the system into a mistake. If it never stops questioning, the experience can become frustrating. Teacher oversight, a sound curriculum and a way to flag misleading responses remain essential.
The role of the teacher does not vanish
Stanford’s wider set of projects includes systems intended to help students find appropriate learning material. Instructors describe considerable human curation of source content. That is an important reminder that an AI teaching interface is not automatically a trustworthy course. Someone still needs to decide which concepts belong in the lesson, what constitutes a valid explanation and how to assess progress.
A teacher can also read the room in ways that are difficult to reproduce through a text conversation. A pupil who is tired, embarrassed or simply unfamiliar with the instructions may need a different intervention from one who has understood the problem but made an arithmetic error. A machine can assist with some forms of personalised practice without fully understanding those human circumstances.
The point is particularly relevant for university students facing independent-study expectations. LiveAIWire has reported on students seeking clearer guidance on appropriate AI use. Rules that only say ‘AI allowed’ or ‘AI banned’ may fail to explain which activities genuinely build competence. LiveAIWire has also looked at how students balance AI empowerment, dependence and trust. A tool that asks questions could fit an educational policy more clearly than one that silently completes assessed work.
What students and families can take from it
Even without a specially designed university assistant, the underlying habit is straightforward. A learner can ask an AI system for a hint, a check on their reasoning or another practice problem instead of a finished solution. They can request that the system withhold the answer until they explain their attempt. That does not make an ordinary chatbot an accredited tutor, but it can change the kind of help being requested.
Parents and students should also be cautious about sharing sensitive school records or identifying personal details simply to obtain more personalised advice. A realistic example or a description of the difficulty may be enough. Accuracy matters too: generated explanations can be convincing even when they contain mistakes, so important conclusions should be checked against trusted course material or the teacher.
The larger lesson is about what efficiency means in education. The shortest route to a correct sentence is not always the shortest route to understanding. Sometimes the productive action is to ask another question, not to type an answer. Stanford’s experiments are a timely illustration of that possibility, with the harder evidence about lasting learning gains still to be established.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.