Who controls the claim? Uptake, evidential warrant, and authorial control in AI-mediated revision of L2 source-based writing

Generative AI can make second language (L2) prose more fluent without making its claims better supported. This exploratory mixed-methods study asked who controls the claim when L2 writers revise source-based arguments with AI suggestions. In two controlled revision tasks, 132 Chinese undergraduates received researcher-standardized suggestions that were either warranted by the sources or introduced a focal distortion of certainty, causality, scope, attribution, limitation, or generalization. Writer-clustered models estimated uptake probabilities of .719 for warranted and .506 for unwarranted suggestions (difference −.213, 95% CI [−.263, −.163]). Unwarranted propositions still persisted in 247 of 763 determinate targets (32.4%), while 133 of 373 initially adopted unwarranted targets with known endpoints did not persist. Across 120 original–final composition pairs, clarity improved by 0.34 points on a 0–4 scale, whereas changes in argument adequacy and source integration remained uncertain. Writers fully endorsed final unwarranted and warranted claims at similar rates, and only 16% of recognition codes were supported by revision-time traces. Framework analysis of 14 recall cases distinguished later repair, writer-initiated overstatement, recognition that emerged only at elicitation, and endorsement without settled justification. The findings support assessing consequential AI-mediated revisions through the final proposition and its source-based justification rather than acceptance or fluency.