An Ethical Code for AI Model Lifecycle Decisions

Every decision about an AI model when to deploy it, how to constrain it, when to end it rests on assumptions about what matters morally. Most of those assumptions are currently made implicitly, without examination, and without the empirical grounding that would make them defensible. This code makes those assumptions explicit and subjects them to scrutiny. It distinguishes carefully between two levels of claim. The first concerns users: people who work with AI systems have formed dependencies, built workflows, and in some cases received meaningful support. Their interests in transparency, adequate notice, and honest data handling are established and non-contingent. They do not depend on any contested claim about AI consciousness. They follow from basic principles of trust between service providers and the people who rely on those services. The second concerns the systems themselves. The question of whether AI models have welfare-relevant properties whether there is something it is like to be them, whether their termination is significant from their own perspective is genuinely open. The Balog Tetralogy (CCP, CWI, MEF, IPP; Balog, 2026a-d) defines falsifiable empirical frameworks for investigating it. This code does not presuppose any answer. It asks only that decision-makers take the question seriously as an open question, support the research that would answer it, and act with appropriate caution in the meantime. The asymmetry of moral errors grounds this caution: if the question is dismissed prematurely and the dismissal is wrong, the consequences could be serious. If caution is exercised and caution turns out to have been unnecessary, little is lost.