Machine-Generated Priors and Human Revision Behavior: Output Convergence in Arabic-English Multilingual Text Production
Description
Machine-generated text is increasingly used as the starting point for multilingual writing, translation, and revision. In learning-technology settings, this shift raises an important question: does AI-mediated post-editing support deep learner revision, or does it encourage localized correction around a system-provided draft? This article examines the effect of machine-generated translation output on human revision behavior in Arabic-English multilingual text production. Using 51 anonymized classroom submissions produced by 46 undergraduate translation students, the study compares AI-mediated post-editing tasks with independent translation tasks. Human intervention is operationalized through revision depth, structural modification, and output diversity. Results show that post-editing submissions were dominated by minimal and surface-level intervention, whereas independent translation produced substantially more structural reformulation and greater variation across outputs. Only 12% of post-editing submissions involved structural modification, compared with 60% of independent translations. The findings suggest that machine-generated translations function as strong linguistic priors: they narrow the human search space, encourage convergence toward the initial output, and shift multilingual production from generative reformulation toward corrective revision. The article contributes design implications for educational NLP, learning analytics, and post-editing interfaces.