Structured AI Integration for Equitable STEM Writing: A Pilot Study of the DISCO Framework

Generative AI now intersects with most student writing. Without structure, students may use it as a shortcut rather than a learning aid. We piloted DISCO, a five-step scaffold (Define, Instruct, Scrutinize, Combine, Optimize), in an upper-division biology course (N = 11). Students wrote a newspaper-style synthesis without AI (baseline), then completed a parallel task after explicit DISCO training with permitted ChatGPT (GPT-4) use. Anonymized essays were double-rated by Biology and English faculty using a seven-criterion rubric; scores were mean-normalized by rater (to reduce fixed stringency) and averaged. Mean scores increased from 55.5% (SD = 12.4) to 66.8% (SD = 15.2), Δ= +11.3 points (95% CI −1.3 to 23.9). Standardized effects were Cohen’s d = 0.60 and Hedges’ g ≈ 0.55 (approx. 95% CI −0.06 to 1.16). Baseline proficiency strongly predicted gains (r= −0.67; 95% CI −0.91 to −0.12): students starting lower improved more. Students reported using AI chiefly for literature triage, summarizing, concept clarification, and outlining rather than drafting. We interpret effects as practically meaningful but exploratory, given small N, fixed task order, and no control group. Instructional materials, anonymized outcomes, and analysis code are provided in a public repository (DOI: 10.5281/zenodo.18217097) to support adoption and replication. Results suggest that structured AI can support STEM writing and may help narrow skill gaps when paired with critical evaluation and reflective practice.