A randomized trial across four sections (140 students) at Wisconsin-La Crosse found students pairing peer study groups with Macro Buddy, a custom Socratic AI tutor, scored highest by the third exam, ahead of solo AI users, group-only students, and a no-AI control.
140 across 4 sections
Students in the randomized trial
4 individual study, group study, individual+AI, group+AI
Experimental conditions
Details
Maturity
Pilot
Promoter
University of Wisconsin-La Crosse (with Appalachian State University)
Period
Fall 2025 (published March 2026)
Keywords
higher education, economics education, AI tutoring, peer learning
Context
Undergraduate macroeconomics students often work through problem sets either alone or in study groups, with instructors unable to be present for every session; instructors at two US universities tested whether a purpose-built AI tutor could substitute for some of that missing guidance.
Objectives
Test whether a custom Socratic-style AI tutor, restricted to course materials, could improve exam performance alone or combined with peer study groups, without simply giving away answers.
Activities
Researchers at the University of Wisconsin-La Crosse and Appalachian State University built 'Macro Buddy', a custom ChatGPT-based tool trained only on the course's own lecture transcripts, slides and homework questions, with general internet access disabled, responding with follow-up questions in a Socratic style rather than direct answers. In a randomized trial, 140 students across four course sections were assigned to one of four conditions: individual study, group study, individual study with Macro Buddy, or group study with Macro Buddy.
Results
Scores on the second exam declined across all four groups. By the third exam, several weeks into the course, the combined condition (peer discussion plus Macro Buddy) scored highest, individual Macro Buddy users outperformed students working alone without AI, and the group-only (no-AI) condition showed the smallest gains.
Conclusions
The authors note the study covered a single course at a single institution, which limits how far the results can be generalized, but deliberately designed the low-cost, no-code tool to be replicable by other instructors.
Implementation
Indicative cost
Low (< €50k) — A custom, low-cost, no-code ChatGPT-based tool built and run by course instructors, deliberately designed to be affordable and replicable by others.
Time to results
Short (< 1 year) — Conducted across a semester (Fall 2025) with results published March 2026; the combined AI-plus-peer-group advantage only emerged by the third of several exams.
Staffing & skills
Course instructors at the University of Wisconsin-La Crosse and Appalachian State University (tool design and deployment)
Conditions for success
Restricting the AI tutor to vetted course materials only, with general internet access disabled
A Socratic response design (follow-up questions, not direct answers) to push independent reasoning
A low-cost, no-code platform deliberately designed for other instructors to replicate
Combining the AI tutor with peer study groups rather than using it as a standalone substitute
Common failure modes
The effect took several weeks to appear — the second exam showed no group advantage and scores declined across all four conditions
Single course, single institution — the authors explicitly caution against overgeneralising
Commonly funded by
Own resources / municipal budget
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
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.
A 13-week controlled study at four Hungarian universities found a ChatGPT-based adaptive tutor using Bayesian Knowledge Tracing nearly doubled programming …