Large language models (LLMs) undergo continuous revision, yet how their moral response tendencies evolve across released editions remains poorly characterized. We present a structured empirical study employing a psychometrically grounded 107-probe instrument covering various moral foundations (e.g., Care, Fairness, Loyalty, Authority, Sanctity) through adapted questionnaire items, ethical dilemma scenarios, value-priority rankings, and meta-ethical probes, administered uniformly across 14 commercial LLMs spanning 2022–2025 from two providers (OpenAI and Anthropic). We introduce the moral drift analysis (MDA), wherein moral drift denotes measurable temporal change in model-expressed value configurations across successive releases, following the established technical usage of drift in machine learning (e.g., data drift) without implying moral degradation or improvement, as a framework that decomposes value change into stance drift and refusal-policy drift. In total, approximately 9,500 model calls reveal a 14-fold asymmetry in mean moral drift between providers: OpenAI models exhibited substantial mean stance drift (Cohen's d = 0.35, exceeding the conventional small-effect threshold of 0.2), characterized by non-monotonic reversal patterns across generations, while Anthropic models maintained temporal stability. Notably, this provider asymmetry was domain-specific, with moral foundation items driving the overall pattern, while ethical dilemma and value priority probes showed mixed or reversed asymmetries. We discuss implications for alignment stability, value lock-in, and the interpretability of moral response patterns in deployed systems, and release code and instruments to support longitudinal auditing of future model updates.
M.Z. Naser (Thu,) studied this question.