Active Learning with Generative AI in Introductory Physics: A Multi-Institutional Study of the P-A.I.R. Framework
Abstract
Generative AI tools have become part of how students study physics, yet unstructured use may reduce students' opportunities to work through problems independently. The Physics AI-Replication (P-A.I.R.) framework structures AI use as a four-step cycle: students identify a challenging problem, obtain an AI-generated explanation and solution, request similar practice problems, and solve them independently before reviewing the AI's solutions. We evaluated P-A.I.R. in 11 algebra- and calculus-based introductory physics sections at four U.S. institutions using a common end-of-semester survey. Most students reported improved conceptual understanding (90.9\%) and found the replicated practice problems helpful (97.7\%), and these two ratings were strongly associated (rs = 0.58). Perceptions of both components, as well as post-engagement confidence (μ= 7.18 on a 1-10 scale), were consistent across institutions. Post-engagement confidence was associated with perceived improvement in conceptual understanding (rs = 0.37) and the helpfulness of replicated practice (rs = 0.35), but not with how often students used AI (rs = 0.02) or with self-reported GPA (rs = 0.13). Thematic analysis of open-ended responses showed that students valued AI primarily for conceptual clarification and structured problem solving (63.0\%), while their main concerns were AI accuracy (34.2\%) and the need for explanations that guide rather than simply provide answers (29.7\%). These findings suggest that how students engage with AI matters more than how often they use it, and that frameworks sequencing AI assistance with independent problem solving offer a practical, transferable approach for physics instruction.
Related
- NuPaD: A Generative AI Framework for Fostering Deep Learning in Subatomic Physics
- Mechanics Cognitive Diagnostic: Testing Fine-Grained Learning Objectives in Introductory Physics
- Calculating a journey to Mars with introductory physics
- Data-driven modeling in the introductory physics laboratory: Scaling analysis and data collapse in the specific heat of water experiment
- Physics-Constrained Conditional Generative Learning for Quantum State and Process Tomography
BibTeX
@misc{paul2026active,
title = {{Active Learning with Generative AI in Introductory Physics: A Multi-Institutional Study of the P-A.I.R. Framework}},
author = {Bilas Paul and Shantanu Chakraborty and Shruti Shrestha and Binod Nainabasti and Bisaj Paul},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.04148},
eprint = {2611.04148},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.04148},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Active Learning with Generative AI in Introductory Physics: A Multi-Institutional Study of the P-A.I.R. Framework
AU - Bilas Paul
AU - Shantanu Chakraborty
AU - Shruti Shrestha
AU - Binod Nainabasti
AU - Bisaj Paul
PY - 2026
DA - 2026/11/03
PB - arXiv
DO - 10.48550/arxiv.2611.04148
UR - https://arxiv.org/abs/2611.04148
AB - Generative AI tools have become part of how students study physics, yet unstructured use may reduce students' opportunities to work through problems independently. The Physics AI-Replication (P-A.I.R.) framework structures AI use as a four-step cycle: students identify a challenging problem, obtain an AI-generated explanation and solution, request similar practice problems, and solve them independently before reviewing the AI's solutions. We evaluated P-A.I.R. in 11 algebra- and calculus-based introductory physics sections at four U.S. institutions using a common end-of-semester survey. Most students reported improved conceptual understanding (90.9\%) and found the replicated practice problems helpful (97.7\%), and these two ratings were strongly associated (r_s = 0.58). Perceptions of both components, as well as post-engagement confidence (μ= 7.18 on a 1-10 scale), were consistent across institutions. Post-engagement confidence was associated with perceived improvement in conceptual understanding (r_s = 0.37) and the helpfulness of replicated practice (r_s = 0.35), but not with how often students used AI (r_s = 0.02) or with self-reported GPA (r_s = 0.13). Thematic analysis of open-ended responses showed that students valued AI primarily for conceptual clarification and structured problem solving (63.0\%), while their main concerns were AI accuracy (34.2\%) and the need for explanations that guide rather than simply provide answers (29.7\%). These findings suggest that how students engage with AI matters more than how often they use it, and that frameworks sequencing AI assistance with independent problem solving offer a practical, transferable approach for physics instruction.
ER -