Machine-Generated Priors and Human Revision Behavior: Output Convergence in Arabic-English Multilingual Text Production

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.