evidoria

← Back to browse

Good practice Imported

Tutor CoPilot — Stanford and FEV Tutor's Human-AI System Lifts Maths Tutoring Outcomes in Title I Schools

United States of America · Woburn · See the United States of America profile

Evidence: Randomised controlled trial Top 5% 80/100 · Ask Evidence Copilot about this practice

A Stanford AI co-pilot gives live, expert-style suggestions to human maths tutors mid-session. A randomised trial with 900 tutors and 1,800 Title I students found a 4-point mastery gain overall, rising to 9 points for less-experienced tutors, at about $20 per tutor a year.

+4 pp
Mastery gain, students with CoPilot-assisted tutors (overall)
+9 pp
Mastery gain, students paired with lower-rated tutors
900
Tutors in trial
1,800
Title I students in trial
$20
Estimated cost per tutor per year
Tutor CoPilot — Stanford and FEV Tutor's Human-AI System Lifts Maths Tutoring Outcomes in Title I Schools

Details

Maturity
Pilot
Promoter
Stanford University (EduNLP Lab) & FEV Tutor
Period
2024
Keywords
K-12 tutoring, mathematics, human-AI collaboration, Title I schools

Context

Tutor CoPilot, developed by researchers at Stanford University's EduNLP Lab with virtual-tutoring provider FEV Tutor, gives human tutors real-time, expert-modelled suggestions -- such as probing questions or ways to unpack a misconception -- while they tutor students live over chat, rather than replacing the tutor.

Objectives

The tool aims to narrow the gap between novice and expert tutors by surfacing, mid-session, the kind of guidance an expert tutor would give, targeted at K-12 maths tutoring in low-income Title I schools.

Activities

A preregistered randomised controlled trial -- the first of its kind for a live human-AI tutoring system -- ran across 900 tutors and 1,800 K-12 students in Title I schools within a US Southern school district's virtual mathematics tutoring programme, comparing tutors randomly assigned access to Tutor CoPilot against a control group without it.

Results

Students whose tutors had access to Tutor CoPilot were 4 percentage points more likely to master the material than the control group; the effect rose to 9 percentage points for students paired with lower-rated tutors, and analysis of more than 350,000 tutoring messages showed CoPilot-assisted tutors asked more probing questions and gave less generic praise.

Conclusions

At an estimated cost of about $20 per tutor per year, the researchers frame Tutor CoPilot as a low-cost way to scale tutoring expertise and narrow quality gaps between tutors, though the trial covers a single district and one subject, and grade-level appropriateness of AI suggestions remains an active area of refinement.

Implementation
Indicative cost
Low (< €50k) — Researchers estimate roughly $20 per tutor per year, framed as a low-cost addition on top of an existing virtual-tutoring programme rather than a new capital investment.
Time to results
Short (< 1 year) — The trial and its published results cover a single tutoring cycle (2024) within one district's programme -- a short, focused evaluation rather than a multi-year rollout.
Staffing & skills
Stanford University EduNLP Lab researchers, FEV Tutor virtual tutors (900 in trial)

Conditions for success

  • a live chat-based tutoring platform that can surface AI suggestions to the tutor mid-session without disrupting the student-facing conversation
  • a preregistered randomised design with a genuine control group, allowing causal attribution of the mastery gain
  • targeting the intervention's benefit analysis at lower-rated tutors, where the largest gains were found, rather than assuming uniform benefit

Common failure modes

  • the trial covers a single US Southern school district and one subject (maths), so generalisation to other subjects or districts is untested
  • grade-level appropriateness of AI suggestions is described as an active area of refinement, not yet solved

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful