Constructing AI safety: a comparative cognitive-discursive analysis of corporate and regulatory discourses

Background The global proliferation of generative artificial intelligence has transformed safety governance into a highly contested sociopolitical battleground, yet existing discourse studies frequently under-theorize the micro-cognitive mechanisms of epistemic reality construction within AI safety standardization. Objective To investigate how corporate and regulatory actors engineer divergent cognitive realities of AI safety through specific micro-grammatical choices. Methods We proposed an integrated Cognitive-Critical Discourse Analysis framework by synthesizing Langacker's cognitive conception of reality with Fairclough's three-dimensional model, and applied it to a comparative corpus of corporate safety frameworks and regulatory policy documents. Results Tech enterprises engineer a reality of techno-pragmatic containment via active voice, high-certainty evidentiality, and immediate projection to naturalize self-regulation. Conversely, regulatory bodies construct a reality of socio-structural vulnerability using distal modals, passive constructions, and deontic mandates to legitimize state-enforced compliance. Conclusions Linguistic normalization is actively deployed to facilitate or contest corporate epistemic hegemony in AI governance. These findings provide empirical references for revising regulatory discourse and enhancing public media literacy against algorithmic exceptionalism.