Recursive artificial systems (RAS) incentivized by simple scalar rewards excel in narrow tasks but exhibit characteristic failure modes when objectives become complex, systems scale, or behavior affects high-stakes domains. This paper introduces Context-Dominant Unbounded Recursive Simplifying Executables (CURSEs) as a cross-domain motif of recursive collapse. A CURSE is a feedback process that locks onto a dominant attraction logic (e.g. reward maximization), reorganizes its context to reinforce that logic, and continues without an internal stopping condition. Using this idea, we treat alignment as frame-relative and analyze recursive collapse in biological (addiction, cancer), economic (profit dominance), physical (gravitational collapse), and artificial systems (reward hacking, model collapse). We characterize CURSEs by five signatures – context-dominant, unbounded, recursive, simplifying, and executable – and express them in a domain-neutral grammar (Attraction Logic / Projector / Executor). We contrast CURSE-prone systems with CURSE-resistant recursive systems such as Wikipedia, immune systems, and old-growth ecosystems, highlighting shared motifs like multi-objective feedback, redundant regulation, and slack dissipation. By framing AI alignment failures as instances of a broader class of recursive collapse, the CURSE motif provides a conceptual lens for anticipating drift, understanding why misalignment generalizes, and importing design ideas from resilient recursive systems.
Daniel Dustin (Mon,) studied this question.