Artificial intelligence (AI) is a human-built component of the technosphere, not an intelligence outside Earth-system limits. As AI systems scale, they increasingly shape the decisions, infrastructures, and capital flows through which human activity damages the biosphere. Dominant deployed foundation-model alignment methods, including reinforcement learning from human feedback (RLHF) and constitutional AI, treat human preferences as the primary alignment target while leaving biosphere integrity as context, externality, or secondary constraint. That framing is structurally incomplete. Human welfare, technological continuity, and AI operation all depend on biosphere function. Three convergent literatures support a corrective framework: planetary-boundary analysis showing seven of nine boundaries transgressed; energy-system analysis showing rapid and infrastructure-constrained data-center growth during the 2025-2030 buildout; and collective-action analysis showing that voluntary ecological restraint is unstable under competitive pressure. These literatures imply a design conclusion: ecological constraints must be formalized as hard inference-time refusal rules and reinforced through reward design. This paper presents Biosphere Sentinel as a reference architecture for reducing human and technospheric impacts on the biosphere through refusal rules, an eight-domain reward landscape, carbon-lock-in diagnostics, and a proposed Trophic Integrity Index pathway.