When AI Writes the Text: AI-Assistance Depth and Recorded Disclosure in University Application Essays

Abstract

Generative artificial intelligence has made assistance with university application essays easier to obtain, but disclosure rules often treat all use as a single category. This study examined whether the absence of AI disclosure varied with the functional depth of assistance. Of 684 applicants recruited through 10 counselling organisations, 617 met the analytic criteria. AI involvement was classified as none, editing only, idea generation or substantive generation using process logs, draft histories and student and counsellor reports, and disclosure was coded from the final application materials. The primary analysis covered the 497 applicants with recorded AI involvement, clustered across 58 counsellors. No disclosure was recorded for 39.9% of editing-only cases, 44.4% of idea-generation cases and 60.4% of substantive-generation cases. After adjustment, idea generation did not differ detectably from editing only, whereas substantive generation carried more than twice the odds of no recorded disclosure relative to editing only (OR = 2.32, 95% CI [1.49, 3.62]) and almost twice the odds relative to idea generation (OR = 1.94, 95% CI [1.31, 2.87]). Intermediary control was not associated with AI disclosure, and no association was found with authorial displacement, ownership, explainability or self-attributed authorship. The pattern is consistent with a functional threshold rather than a gradient. Universities should define material assistance by function and record which disclosure route each applicant was offered.