Stanford's Tutor CoPilot lifts math mastery 4 percentage points

AI system

System: Tutor CoPilot, integrated into FEV Tutor. [1] [2]
Model: GPT-4. [1] [2]
Release: The author identifies GPT-4 but does not give the trial’s release identifier. [1] [2]
Provider: Stanford researchers developed Tutor CoPilot with FEV Tutor. OpenAI supplied GPT-4. [1] [2]
Role: GPT-4 proposed responses that human tutors could use during lessons. [1] [2]
Configuration: Bridge supplied teaching strategies. The system removed names and sent at most the 10 latest chat messages, the lesson topic, and the requested strategy. [1] [2]
Inputs and outputs: The model used recent messages, the lesson topic, and a teaching strategy to suggest a tutor response. [1] [2]
Autonomy: Tutors chose, edited, regenerated, or ignored the suggestions before responding to students. [1] [2]
Workflow: A language model prompted with a taxonomy of expert tutoring moves rather than answers. Integration into a live virtual-tutoring platform where tutors can request suggestions mid-session. Tutor-level randomization (900 tutors) with preregistered analysis.
Human oversight: Not assessed
Relation to AI: The model advised the human tutor and never taught the student directly. The measured change runs through tutor behavior (more prompting for explanations, fewer given-away solutions). Outside researchers quoted by MIT Technology Review note that the mathematics was relatively basic.
AI contribution: The model advised the human tutor and never taught the student directly. The measured change runs through tutor behavior (more prompting for explanations, fewer given-away solutions). Outside researchers quoted by MIT Technology Review note that the mathematics was relatively basic.

Benefits

Outcome identifiers: Not assessed
Benefit: students of treated tutors were 4 percentage points more likely to master topics
Category: Improved Learning
Status: Not assessed
Beneficiary groups: People. Organizations.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Measurement

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: students of treated tutors were 4 percentage points more likely to master topics
Comparison: students of control-group tutors on the same platform who did not have access to Tutor CoPilot
Results: Reference: 62% student passing rate on exit tickets (control tutors). Result: 66% passing rate (tutors with Tutor CoPilot), 4 percentage points higher.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Money

Kind of value: Not assessed
Amount: Not assessed
Currency and prices: Not assessed
Period and unit: Not assessed
Realization: Not assessed
Recipients and payers: Not assessed
Gross or net: Not assessed
Included costs: Not assessed
Comparison: Not assessed
Method: Not assessed
Return and payback: Not assessed

People and places

Participants: Developer: Stanford University (Wang, Ribeiro, Robinson, Loeb, Demszky). Operator: FEV Tutor with a large southern U.S. school district (nine schools).
Jurisdictions: Developer: United States. Operator: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Follow-up

Conditions for replication: Human tutors willing to use suggestions in real time. Basic-level mathematics content (outside researchers' caveat). Students below grade level in under-served schools, the population studied.
Replication attempts: Not assessed
Later findings: The preprint was revised again on 10 October 2026, but no later deployment, replication or publication of the trial was found.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Evidence

Source numbers: AIBD-0216-source-0001–0005
Citations: Numbered references below.
Language: Not assessed
Source relationships: Developer, Independent
Original evidence: The trial paper supplies the results. The author’s demonstration notebook identifies GPT-4. Earlier and later versions of the paper are one source. [1] [2]
Access and versions: The paper and a dated copy of the author’s demonstration notebook were accessible. [1] [2]
Supported findings: Not assessed
Evidence links: Not assessed
Evidence strength: Not assessed
Missing evidence: Not assessed

Grades

Assessed benefits: People. Organizations.
Depth: Not assessed
Reasons: People: Students of tutors with the tool were 4 percentage points more likely to master topics, 9 points for lower-rated tutors (preprint abstract). Organizations: The tutoring provider's lower-rated tutors improved most at about $20 per tutor annually (preprint abstract).
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Entry details

Kind: Not assessed
Related entries: Not assessed
Field: Education. K-12 virtual math tutoring (grades 3-8).
Purpose: Tutor CoPilot was built to scale expert tutoring guidance to novice tutors in real time. Higher student mastery is that purpose.
Dates: Date basis unassessed: 2024 (precision unassessed). Updated: 25 October 2026.
Stage: Not assessed
Version: SHA-256: ba5e099c4142…

References

Other identifiers: Not assessed
Category mappings: Not assessed
Brief: Not assessed
Revisions: 6 revisions. Latest: 25 October 2026.

Sources

Preprint Developer

Wang RE et al. Tutor CoPilot, arXiv:2410.03017v2 (study start and partner details).

Documents the deployed workflow and the tutor’s control over responses. §3, Tutor CoPilot

Original research or documentation from the system’s researchers. This supports identity and methods, not independent verification of benefits.
Other Developer

Rose E. Wang, Tutor_CoPilot_Demo.ipynb, commit 25c6eab7e416614f2033358bceddd30a4d5ca616, October 28, 2024.

The author explicitly identifies GPT-4 as the model used in the original work. Demo code is not proof of a dated production model release. generate(prompt) cell, model argument and adjacent comment

Original research or documentation from the system’s researchers. This supports identity and methods, not independent verification of benefits.
News report Independent author

Langreo L (31 Oct 2024). What Happens When an AI Assistant Helps the Tutor, Instead of the Student. Education Week (quotes Susanna Loeb, Stanford).
News report Independent author

Toppo G (7 Oct 2024). Study: AI-Assisted Tutoring Boosts Students' Math Skills. The 74 (quotes Rose E. Wang, Susanna Loeb, Amanda Bickerstaff, Benjamin Riley, Dan Meyer).
News report Independent author

Williams R (28 Oct 2024). This AI system makes human tutors better at teaching children math. MIT Technology Review (quotes Rose Wang, Simon Frieder of Oxford, Mina Lee of Chicago).