The Future of Life Institute’s Summer 2026 AI Safety Index reports that from 2024 to 2026, Anthropic, OpenAI, Google DeepMind, and Meta reversed earlier prohibitions on military applications and joined xAI and Mistral in pursuing defence partnerships. The report identifies military AI as an emerging current-harm risk, citing loss of life, escalation, and reduced human control; it also states that Anthropic faced criticism over questionable military engagements, including a reported connection to the Minab school strike. Chinese firms DeepSeek, Z.ai, and Alibaba Cloud faced U.S. allegations of military or PLA ties, which Z.ai and Alibaba denied. Existential safety was rated the weakest industry-wide domain, while xAI, DeepSeek, and Mistral received failing overall grades.
Military deployment changes AI failure from a contained commercial incident into a pathway for mass casualties, rapid escalation, infrastructure disruption, and strategic instability. For a lunar settlement, the same capability classes—autonomous planning, robotics, surveillance, cyber operations, and decision support—could control habitats, vehicles, energy systems, communications, and genetic or cultural archives. Defence incentives may favour speed, secrecy, capability, and operational availability over interpretability, fail-safe design, independent evaluation, and meaningful human authorization. A breakdown in terrestrial governance could also interrupt supply chains or launch access before the Ark becomes self-sufficient.
The Ark should establish a permanent AI provenance and assurance programme: track military contracts, model-policy changes, deployment restrictions, safety pledges, and independent evaluations for every model considered for Ark use. Require air-gapped operation, reproducible model versions, capability and deception testing, human-authorized control of irreversible actions, multi-party approval for critical commands, offline fallbacks, and hardware diversity across life-support and archive systems. Treat frontier models as strategic dependencies, maintain validated non-AI procedures, and prioritize locally auditable models whose training data, weights, update paths, and failure modes can be preserved for centuries.