Machine Behavior in Relational Moral Dilemmas: Moral Rightness, Predicted Human Behavior, and Model Decisions

The study's most underreported finding is that models maintain a third, distinct behavioral mode when asked how they themselves would act, separate from both the prescriptive ethics framing and the human-prediction framing. That three-way split suggests LLMs aren't just inconsistent; they're running something closer to parallel, context-sensitive reasoning tracks that can produce contradictory outputs from nearly identical prompts.

This connects to the hallucination work covered here around the same period: 'When Prompts Override Vision' showed that textual priors in prompts systematically distort model outputs in vision-language systems. The same mechanism appears to be operating in moral reasoning tasks, where framing a question as prescriptive versus predictive functions like a different prompt prior, pulling the model toward different output distributions. Neither paper is about ethics per se; both are about how sensitive model behavior is to surface-level input variation. The RedirectQA coverage on entity surface forms reinforces the same underlying point: LLM outputs are far more contingent on how a question is posed than on any stable internal representation of the answer.

Watch whether follow-up work tests these three framings against models fine-tuned on RLHF with explicit human-preference data. If the prescriptive-predictive gap narrows in those models but the self-decision track diverges further, that would confirm the gap is a training-objective artifact rather than a fundamental property of scale.