AI-generated text detectors often achieve high in-domain accuracy but remain sensitive to changes in domain, generator, and writing style. This limitation partly arises because conventional detectors rely predominantly on latent language-model representations and fixed decision boundaries, without explicitly characterizing sample-level structural and semantic variation. We introduce S²E-AL, an end-to-end score-aware structure–semantic entropy adversarial learning framework for AI-generated text detection. S²E-AL encodes local dependency, global contextual, and semantic entropy features into a structure–semantic representation. A gradient-reversal source classifier discourages this representation from retaining source-specific information, while a score-aware threshold estimator conditions on both the entropy representation and a detached base-detector logit to produce a sample-specific logit adjustment. The adjusted logit is used to estimate the final probability that a text is AI-generated. Across HC3, HC3_Plus, LLMTrace, RAID, and XSum, the score-aware model achieved a macro Accuracy of 90.90% and a macro F1-score of 91.42%. Its performance remained lower on XSum, indicating that substantial domain and stylistic shifts remain challenging. These results suggest that jointly modeling structure–semantic variation and detector confidence can improve the overall performance–stability trade-off of AI-generated text detection across heterogeneous test distributions.