Introduction Generative artificial intelligence (GenAI) can adapt English for Academic Purposes (EAP) reading materials by rewriting passages, adding support, or combining both. This study examined whether proficiency-sensitive GenAI adaptation chiefly changed passage-level structural complexity or text-embedded functional support while preserving academic fidelity. Methods A role-prompted workflow combined barrier analysis, adaptation, fidelity checking, and validation. In a 3 × 3 between-subjects design ( N = 135; n = 15 per cell), proficiency level (low, intermediate, and high) was crossed with material condition (original, unified-AI, and differentiated-AI). Three EAP instructors evaluated 15 anonymized material versions. Automated structural-complexity indicators, leave-one-out discriminant analysis, within-proficiency planned contrasts, and omnibus outcome models were used to assess material and learner outcomes. Results The instructors rated the 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 ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M1"> mml:mrow mml:msubsup mml:mrow mml:miη</mml:mi> </mml:mrow> mml:mrow mml:mip</mml:mi> </mml:mrow> mml:mrow mml:mn2</mml:mn> </mml:mrow> </mml:msubsup> mml:mo=</mml:mo> mml:mn0</mml:mn> mml:mo.</mml:mo> mml:mn043</mml:mn> </mml:mrow> </mml:math> ), and leave-one-out discriminant analysis classified the intended proficiency labels at 11.1%, below the 33.3% balanced-task benchmark. Material differences were clearest in functional support, including glosses, sentence unpacking, rhetorical cues, claim-evidence notes, and critical prompts. 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 &lt; 0.001, and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M2"> mml:mrow mml:msubsup mml:mrow mml:miη</mml:mi> </mml:mrow> mml:mrow mml:mip</mml:mi> </mml:mrow> mml:mrow mml:mn2</mml:mn> </mml:mrow> </mml:msubsup> mml:mo=</mml:mo> mml:mn0</mml:mn> mml:mo.</mml:mo> mml:mn191</mml:mn> </mml:mrow> </mml:math> . Supplementary process estimates were descriptive because the process indicators and outcomes were collected in the same session. Immediate unsupported application did not differ by material condition or interaction. Discussion The findings locate GenAI-supported differentiation primarily in proficiency-specific support-layer design rather than broad changes in passage-level structural complexity. The strongest learner evidence came from comparisons conducted within the same proficiency level.