This paper turns safety from a neutral technical promise into a field of corporate speech. Generative AI companies do not merely describe safety; they manufacture its public grammar. Safety becomes a discursive instrument through which authority, responsibility and legitimacy are arranged before regulation can fully intervene. The crucial methodological move is reading company documents as artifacts of power, refusing the innocent surface of public-facing language. Corporate statements are not afterthoughts attached to technology; they are infrastructures of permission. They tell users, legislators, researchers and markets what counts as risk, who is qualified to name risk, and which remedies appear reasonable. The paper displaces safety from engineering to politics. The corporate idiom presents AI as inevitable, transformative and broadly beneficial, while risk appears as continuous experimentation, anticipatory alignment and responsible deployment. This vocabulary distributes agency: companies become stewards of a future they are accelerating; users become participants in an unfolding safety process; affected communities are often invited into a narrow procedural role after the main technological direction has been decided. The paper therefore reads safety as a soft regime of governance: a way of producing consent before accountability.