Generative artificial intelligence is propelling accessible digital publishing beyond mere format conversion toward semantic generation,interactive Q&A,and personalized adaptation.Yet existing discourses fall short of a systematic examination of how risks diverge across publication types,disability types,technical workflows and normative evaluation.This study constructed a four-dimensional scenario-based analytical model cross-cutting publication scenarios,disability types,technical workflows and normative evaluation based upon the"copyright jurisprudence-data ethics-content value"framework.Findings revealed that literary works were predominantly susceptible to stylistic drift and voice rights and interests infringement,textbooks,popular science,medical,and legal publications faced core risks of factual and causal distortions,while news and emergency information were vulnerable to error amplification under time pressure.Visually,hearing,and cognitively impaired users exhibited different levels of vulnerability to omissions in image descriptions,deviations in captioning and sign language,and mistakes in easy-to-read processing,respectively.Although current legal regimes have formed a rudimentary institutional chain,they still lack process-oriented risk stratification,perceptible compliance labeling,participatory testing by persons with disabilities,and auditable accountability trails.To address these gaps,this study proposed an enforceable regulatory scheme encompassing risk gradation,review procedures,principal accountabilities,data controls,copyright boundaries,and post-release monitoring,with clearly delineated four-tier review standards,lifecycle protection rules for sensitive personal information,and fair-use parameters for accessible format editions.The ultimate objective is to steer accessible publishing from mere"accessibility in form"toward genuine"comprehensibility,controllability,and trustworthiness in substance".