Ask an AI whether someone should do something ethically fraught, then ask whether they should not — and you may get two opposite answers.
Researchers audited 16 language models across 14 ethical dilemmas by pairing each question as both a prescription ("They should X") and a prohibition ("They should not X"). A model with a coherent moral stance should give the same underlying judgment either way. Many don't. Small open-weight models in the 1-4 billion parameter range endorsed a proposed action only 24% of the time under affirmative framing, but up to 100% of the time when the same action was framed as a prohibition — a swing of as much as 76 percentage points. Larger commercial models held up better but still shifted: cross-model agreement fell from 73% on straightforward affirmative questions to 59% once simple negation was introduced. Human coders confirmed the instability is real, though they also found that binary agree/disagree scoring overstates it.
The finding matters because AI systems are increasingly consulted on decisions with real consequences — medical triage, legal guidance, content moderation — and a model whose moral output depends on sentence construction rather than the underlying situation is not expressing a stable position at all. The paper proposes a new metric, the Negation Sensitivity Index, to measure stance stability directly, arguing that single-phrasing audits can misreport what a model actually "believes" about an ethical question.
The irony is that using another AI to judge the results doesn't help: the study found LLM-based evaluators silently collapse ambiguous responses and reproduce the same forced-choice bias being measured, which means the tools most commonly used to scale AI evaluation may be hiding exactly the problem this research is trying to expose.