Back to Bigger Picture

BIGGER PICTURE Learning & work

The tutor gets a coach at the difficult moment

A pupil gives a wrong answer. Tutor CoPilot helps the human tutor choose a useful question or hint, turning AI into support for better teaching rather than an answer dispenser.

AI-generated illustration: A human tutor receiving discreet hints while helping a learner.
A human tutor receiving discreet hints while helping a learner. Conceptual illustration; not a photograph or an exact depiction of the real project.AI-generated illustration of the idea, not a photograph of the actual people, equipment or event.
Math-topic mastery
+4 percentage points
With lower-rated tutors
+9 percentage points
Who receives the suggestions?
The human tutor

How the idea works

  1. Recognize a teaching impasse

    The tutor requests help during a specific exchange.

  2. Suggest teaching strategies

    Topic and conversation produce questions, hints or explanations.

  3. Let the tutor adapt

    A person chooses the response and checks understanding.

What to say after a mistake

A good tutor knows when to ask a question, offer a hint or try another explanation. Less experienced tutors may instead reveal the answer or give praise that does not address the misunderstanding. Tutor CoPilot places guidance at that precise decision point, inside a live tutoring conversation rather than in a separate training course.

Sources: [2]

Three suggestions, one responsible person

The tutor activates the tool when needed. It uses the lesson topic and de-identified conversation to suggest several teaching responses informed by expert reasoning. The tutor can select a strategy, edit a suggestion or ignore it. This preserves something a general chatbot lacks: the educator’s knowledge of this learner and the immediate teaching context.

Sources: [2][3]

A measurable gain in real tutoring

The updated working paper reports a randomized trial involving more than 700 tutors and 1,000 students from underserved communities. Students whose tutors had access were four percentage points more likely to master the session’s mathematics topic. The improvement reached nine points for lower-rated tutors, suggesting that timely guidance can narrow differences in instructional quality.

Sources: [1]

Better questions are the important product

Analysis of tutoring messages found more probing questions and less generic praise. But tutors also reported suggestions that were unsuitable for the pupil’s grade level. The measured outcome was lesson mastery, not a demonstrated year-long attainment gain. The study supports coaching the adult in a specific setting; it does not establish that unmonitored AI tutoring works equally well.

Sources: [1][3]

The original sources

  1. EdWorkingPapers / Annenberg InstituteTutor CoPilot: updated working-paper record
  2. Stanford research teamTutor CoPilot: November 2025 working paper
  3. Research authors / GitHubTutor CoPilot demonstration and source code

What to explore next

Six weeks of guided practice changed what pupils could doThe scan that helps a stroke patient reach the right hospital
Explore all 99 applications