Cognitive Predictors of AI-Related Textual Change in Upper-Intermediate L2 Russian Writing: Working Memory, Attention Control, and Critical Thinking

Generative artificial intelligence (AI) can alter final second-language (L2) texts, yet cognitive differences may shape which changes emerge. This exploratory study examined 38 upper-intermediate learners of Russian who each produced one composition with AI assistance and one without it. Working memory, attention control, and critical thinking were measured continuously, and 11 textual outcomes were calculated from paired compositions. AI-assisted texts showed higher lemma-based lexical-diversity values, longer words, more connectives per 100 words, and substantially fewer first-person markers, while total word count, sentence measures, and formulaic-expression rate did not differ in the full sample. The higher POS-based content-word density was attenuated after first-person forms were excluded, suggesting a contribution from the change in person marking. Exploratory models suggested negative working-memory associations with lexical-diversity and connective-rate change, and negative attention-control associations with lexical-diversity change. The positive working-memory coefficient for word-count change was sensitive to HC3 estimation and leave-one-out refitting. Critical thinking showed no clear independent association in the primary change-score models, while positive coefficients for MATTR and HD-D appeared in the outcome-conditioned models. Under this supervised, student-mediated procedure, AI assistance was associated with lexical, discourse-marking, and stance differences in completed L2 texts.