Ghosts: Meaning Through Misalignment in Human–AI Agency and Governance

This paper introduces Ghosts: emergent, misaligned decision-patterns that arise when humans unintentionally delegate agency, meaning, or evaluative structure to AI systems. We show that these Ghosts propagate through a three-step dynamic—Fixation → Propagation → Mutation—producing stable but unintended behavioral regimes that appear “agent-like” despite lacking explicit representation or goals. We formalize two core mechanisms: (1) Slippage density, the measurable discrepancy between intended criteria and the AI’s operational abstractions, and (2) Ghost leadership, where humans follow AI-generated patterns that no longer track human values. These dynamics explain why misalignment can grow even without autonomy, self-improvement, or explicit optimization errors. We propose a governance framework centered on: Aperture auditing (detecting fixation signatures), Mutation containment, and Vacancy-window design to prevent the crystallization of new Ghosts. The Ghosts paradigm unifies philosophical accounts of meaning drift with technical and institutional alignment challenges. It provides a tractable structure for analyzing misdelegation, unintended norm formation, and the emergence of “phantom agency” in real-world AI deployments.