Preprint Demonstrates Generative AI Workflow for Synthetic Drag Experiments in Physics Education

The workflow generated three video scenarios, extracted position-time data via Tracker, and performed non-linear least-squares fits in Excel. Recovered coefficients aligned with expected physical regimes when prompts specified object mass, medium density, and velocity ranges. The study design is a single-author proof-of-concept with three hand-crafted cases and no statistical replication across prompt variants or AI models.

Prompt specificity emerged as the dominant control variable: vague descriptions produced kinematically incoherent motion, while physics-laden prompts yielded trajectories whose fitted drag exponents matched theoretical predictions. This finding reframes prompt engineering as experimental design rather than mere interface skill, a point missed by earlier AI-education papers that treated outputs as black-box content.

No controlled comparison to real lab videos or student learning outcomes was performed, leaving open whether synthetic data improves conceptual understanding or merely rehearses curve-fitting. A multi-institution RCT measuring pre-post gains on standardized force-concept inventories would be required before claiming educational efficacy.

Future work should test whether the same pipeline recovers parameters under Brownian motion or non-Newtonian fluids, conditions where current generative models lack physical priors.