The AI that helps a tutor ask a better question

The decisive question in tutoring is not always, “What is the answer?” When a student makes a mistake in subtraction or fractions, the next step can reveal whether they understand the process or have merely memorized a rule. Giving a hint that is too direct finishes the exercise; asking the right question can help build learning.

Tutor CoPilot, developed by Stanford researchers, applies artificial intelligence to that moment. It is not a chatbot that replaces the tutor or an assistant that speaks directly to the student. It is a tool for the person delivering tutoring: when activated after an error, it proposes guided strategies and questions that the tutor can use, adapt, or ignore.

The design was tested in an online mathematics tutoring programme run by FEV Tutor with a U.S. Southern school district. The 2024 study was a preregistered randomized trial and took place in Title I schools, which receive federal support because they serve students from low-income families.

What it does in a session

The tutor remains in conversation with the student. If the student is stuck, Tutor CoPilot does not provide a finished solution. It may suggest that the tutor ask the student to identify the numbers in a subtraction problem or represent the items before calculating. The aim is to bring some of an expert educator’s pedagogical reasoning into the moment when a less-experienced tutor needs to respond.

The tool uses language models, but its guidance was designed from experienced educators’ reasoning processes. It is tutor-facing and offers several strategies; the tutor keeps the decision of when to activate it, which suggestion to use, and how to adapt it to the student’s context. That separation matters: an AI suggestion does not by itself know a class’s full history, a student’s actual level, or a particular support need.

What was measured

The 2024 version describes a trial with 900 tutors and 1,800 elementary and secondary students. Students whose tutors had access to Tutor CoPilot were four percentage points more likely to master the assessed topic: exit-ticket pass rates rose from 62% to 66%. Among tutors initially rated lower, the improvement reached nine percentage points, from 56% to 65%.

The study also analysed more than 550,000 tutoring messages. Tutors with access to the tool used strategies associated with deeper learning more often, such as questions that guide reasoning, and were less likely simply to give away the answer. The estimated cost, based on trial usage patterns, was about US$20 per tutor per year.

Those numbers explain the case’s interest. AI is not trying to replace educational interaction; it is trying to give a live tutor more resources to respond well to an error. The larger gain appeared where access to expert guidance is often most limited: among tutors with lower initial ratings.

Why human control is part of the outcome

In education, a fluent answer is not enough. It needs to be correct, age-appropriate, consistent with the curriculum, and useful for helping a student think. Tutors themselves identified an important limit: some Tutor CoPilot suggestions were not well matched to the grade level. In early tests, the system could also take more than 30 seconds to respond, an awkward delay in a live conversation.

Both issues show why a tutor cannot simply copy an output. They need to check whether a question makes sense, rephrase it, and decide whether it is the right moment. The tool offers support; pedagogical responsibility remains human.

Privacy also needs specific design. Stanford said the system automatically de-identified student and tutor names and limited information sent to external model services. That reduces exposure, but it does not remove the duty of districts to review contracts, data retention, security, and consent.

What it does not yet establish

The trial measures math-topic mastery in one programme, one district, and one delivery mode. It does not by itself show that every chatbot improves learning, that effects persist over time, or that they transfer unchanged to reading, science, or in-person settings. It also does not prove that the same cost holds when every system scales it.

The research itself notes that off-the-shelf language models do not necessarily behave well in real educational interactions. This case involved pedagogical design, an existing tutoring platform, and a controlled trial. Removing those elements and offering a general chat interface would be a different case with different risks.

A lesson for schools

Tutor CoPilot offers a restrained way to use AI: improve a human intervention that already exists. Rather than asking a machine to educate alone, it gives it a bounded job: suggest the next question when a tutor needs help.

For a school or district, evaluation must go beyond whether a tool seems impressive. It should measure learning, suggestion quality, actual use, response times, errors by grade or student group, and data protection. Educational AI is not measured by how many answers it generates, but by whether it helps a student reason independently.