Answers are grounded in this paper and cite their sources.
Generative artificial intelligence can accelerate drafting, but it can also make unsupported source use harder to notice because fluent prose may appear warranted before the writer has checked the evidence. This quantitatively driven explanatory sequential mixed-methods randomized intervention study tested whether a small amount of source-focused transfer friction in an artificial-intelligence writing interface supported learning rather than only drafting efficiency. Ninety-six undergraduate English majors in a compulsory second-year academic writing course at a public university in Jordan were individually randomized to one of two supervised laboratory conditions for four weeks: an access-matched standard chatbot interface or SourceTrace, a browser-based interface that required students to classify rhetorical move, judge evidential status, generate a counter-position, and justify source use before machine-generated text could be transferred into the draft. Both conditions used the same GPT-4o model snapshot, source packets, prompt-function menu, generation settings, and supervision procedures. All students completed an unaided pretest essay, four assisted source-based writing sessions, an unaided posttest transfer essay, and a source-evaluation task; process logs and stimulated-recall interviews with a maximum-variation subsample of 24 students supported explanatory integration. Controlling for pretest performance, course section, and placement score, students assigned to SourceTrace outperformed students assigned to StandardChat on posttest source-grounded argument quality by 1.38 points on a 20-point rubric. SourceTrace was also associated with stronger source-evaluation accuracy, lower unverified uptake, higher source-justified uptake as the behavioral proxy for calibrated reliance used in this study, and deeper revision. The findings support a bounded design claim: in supervised L2 source-based argument writing, a bundled transfer-gating interface can preserve generative support while making source accountability harder to bypass.
No takes yet. Share an insight, caveat, or question.
Aldamen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: