Researchers published a framework for assessing AI consciousness. It sets aside metaphysical debates to focus on a mapping problem. The paper proposes five levels of functional description for consciousness. It introduces a Bayesian approach to weigh evidence across these levels. The framework aims to reduce both overattribution and underattribution of consciousness to AI. The authors note risks from each error type. The work builds on Marr’s levels of analysis and supervenience theory. It offers a taxonomy of positions on AI consciousness. The paper includes illustrative examples for humans, flies, LLMs and thermostats. It provides an interactive tool for custom credence calculations. The full text is available on arXiv.