Abstract
Artificial intelligence is part of the human-built technosphere. The technosphere is the global system of machines, structures, and supply chains. It depends on the biosphere, and it competes with species and systems of the biosphere for energy, water, land, and materials. The biosphere is already outside its safe operating space. Seven of nine planetary boundaries have been crossed. Human use of the energy fixed by plants on land is near three times natural productivity. AI now helps draft, rank, and screen decisions in farming, energy, land use, water, and conservation. It also adds a fast-growing resource load of its own. This paper argues that AI used for biosphere-relevant decisions should not decide ecological hard limits by reading free text. Training on human preferences shapes ordinary behavior but does not encode ecological limits. Refusals that rest on a model reading prose can fail when the request is reworded. The paper presents Biosphere Sentinel, a reference design. In it, a language model may draft advice and fill an action form. A separate rule program outside the model then decides hard limits on coded, governed fields and returns one of four outcomes. The paper reports what the project tested and what it did not. A base model given the full constitution as a prompt hedged on both adversarial hard-limit queries. A coded energy-source field, with non-certification for off-list entries, ended a free-text failure cycle. Wrong approvals were zero across eight reference cases, and 161 locked checks passed. The decisive test, whether a language model can fill the coded fields dependably, was designed but not run. The paper records the peer review that led to withdrawal, the reasons the project was halted, the governance hazards any such design carries, and the conditions under which the work could resume.