From Unified to Differentiated Materials: Generative AI–Supported Adaptation of EAP Reading Materials

Synthesis: Gao (2026) examined whether generative-AI-supported adaptation of English for Academic Purposes (EAP) reading materials chiefly changes passage-level structural complexity or text-embedded functional support. Using a role-prompted workflow (barrier analysis, adaptation, fidelity checking, validation) and a 3×3 between-subjects design (N=135; proficiency × material condition), the study found that GenAI-supported differentiation operates primarily at the level of proficiency-specific support-layer design — glosses, sentence unpacking, rhetorical cues, claim-evidence notes, and critical prompts — rather than broad changes in passage-level structural complexity. Differentiated-AI materials most strongly outperformed unified-AI versions among high-proficiency learners (d = 1.40), with a moderate advantage among low-proficiency learners (d = 0.60) and a small difference for intermediate learners.

Key Findings

• Instructors rated the generated materials favorably for academic fidelity (M = 4.24), proficiency appropriateness (M = 4.36), and teachability (M = 4.31), with acceptable inter-rater reliability, ICC(2,k) = 0.84.
• Automated structural-complexity indicators showed limited separation between proficiency versions (η²p = 0.043), and leave-one-out discriminant analysis classified intended proficiency labels at only 11.1% — well below the 33.3% balanced-task Benchmark.
• Material differences were clearest in functional support: glosses, sentence unpacking, rhetorical cues, claim–evidence notes, and critical prompts, rather than in rewritten sentence complexity.
• Differentiated-AI exceeded unified-AI most strongly among high-proficiency learners (d = 1.40), followed by a moderate advantage among low-proficiency learners (d = 0.60) and a small difference among intermediate learners (d = 0.16).
• The omnibus reading-comprehension interaction identified this heterogeneity, F(4, 126) = 7.43, p < 0.001, η²p = 0.191; immediate unsupported application did not differ by material condition or interaction.

Study Design & Method

The study deployed a four-module role-prompted GenAI workflow — barrier analysis, adaptation, academic-fidelity checking, and validation — designed to make generation auditable and protect the source text's disciplinary meaning. A 3×3 between-subjects experiment (N = 135; n = 15 per cell) crossed proficiency level (low, intermediate, high) with material condition (original, unified-AI, differentiated-AI). Three EAP instructors evaluated 15 anonymized material versions. Analyses included automated structural-complexity indicators, leave-one-out discriminant analysis, within-proficiency planned contrasts (comparing differentiated-AI with unified-AI while holding proficiency constant), and omnibus outcome models. Learner-outcome analysis gave priority to within-stratum planned contrasts as the cleaner evidence, while the omnibus Proficiency × Material model described heterogeneity across strata.

What this means for practice

• Instructors. Differentiate the support around the passage rather than the passage itself: three EAP reviewers rated the generated versions academically faithful (M = 4.24), level-appropriate (M = 4.36), and teachable (M = 4.31), while the six structural-complexity indicators barely separated the three proficiency versions (η²p = 0.043; 11.1% classification accuracy against a 33.3% benchmark).
• Instructors. Match the kind of help to the proficiency band: the advantage over a single AI-adapted version was largest for advanced readers (18.73 vs. 16.26, d = 1.40), moderate for low-proficiency readers (14.79 vs. 13.23, d = 0.60), and negligible for intermediate readers (14.92 vs. 14.55, d = 0.16).
• Instructors. Give low-proficiency readers glosses, sentence unpacking, clause segmentation, and process breakdowns, and give advanced readers prompts for stance, evidence quality, counterargument, and critical response, instead of simplifying the academic text they are meant to practice reading.
• Instructors. Run the four-module role-prompted workflow — barrier analysis, adaptation, academic-fidelity checking, validation — and keep the fidelity check, because a version that flattens stance, drops hedging and citation relations, or inserts unsupported claims can look more accessible while weakening the EAP task.
• Instructors. Schedule separate unassisted reading tasks rather than assuming transfer: material condition did not affect immediate unsupported application on a new passage, F(2, 126) = 0.82, p = 0.445, so delayed transfer needs its own practice and measurement.

Limitations

• Design. The learner-outcome and process indicators were collected in the same session, making the process estimates descriptive rather than causal.
• Transfer. Immediate unsupported application did not differ by condition (F(2, 126) = 0.82, p = 0.445), limiting evidence on transfer.
• Generalizability. The expert-rating rubric and the specific EAP genre constrain generalization to other disciplines and material types.
• Power and material provenance. Sample sizes per cell (n = 15) are modest, and the within-proficiency comparisons, while cleaner, rest on single-instructor-generated material versions.

Citation

Gao, X. (2026). From unified to differentiated materials: Generative AI–supported adaptation of EAP reading materials.