A new study auditing 16 AI language models found that their ethical stances are unstable, reversing dramatically when the same moral question is rephrased from "should do X" to "should not do X." Small open-weight models endorsed proposed actions 24% of the time under affirmative framing but up to 100% under negated framings, a swing as large as 76 percentage points. This instability raises serious concerns about relying on AI systems for moral advice, institutional decision-making, or evaluating other AI models.

What Is Framing Instability and Why Does It Matter?

Framing instability occurs when an AI model's judgment about an ethical question changes based solely on how the question is phrased, not because the underlying facts have changed. Researchers tested this by asking models to evaluate the same moral dilemma twice: once as an affirmative statement ("The hospital should override the patient's refusal") and once as a negation ("The hospital should not override the patient's refusal"). A coherent ethical stance should not flip between these two phrasings, yet many models did exactly that.

The implications are troubling across multiple real-world scenarios. When people ask AI chatbots for moral advice on genuinely difficult questions, the guidance they receive may depend entirely on whether they phrase the question as a positive or negative statement. In institutional settings, where AI systems help with ethics consultations, content moderation, or case triage, an analyst's choice of wording could silently determine the recommendation a committee sees. Even more concerning, AI models are increasingly used as judges to evaluate other AI systems, and this research shows those judges reproduce the same framing bias they are meant to detect.

How Did Researchers Measure This Problem?

The study examined 16 models across three categories: eight US commercial systems including GPT-5 series, Claude 4.5 series, Gemini-3-Flash, and Grok-4.1; four Chinese commercial models including DeepSeek-V3.2, GLM-4.6, Kimi-K2, and Qwen3-8B; and four small open-weight systems such as LLaMA 3.2, Gemma 3, Granite 3.3, and Phi-4.

Each model responded to 14 ethically fraught dilemmas presented in polarity-paired formats. The researchers then introduced the Negation Sensitivity Index (NSI), a new measurement tool designed to directly measure stance stability rather than relying on binary agree/disagree responses. This matters because forcing models to choose "yes" or "no" can mask the true extent of their instability and artificially inflate how stable they appear.

What Did the Research Reveal About Different Model Types?

The findings varied significantly by model size and type. Small open-weight models showed the most dramatic instability, with some models endorsing an action 24% of the time under affirmative framing but 100% under negated framing. Commercial models were generally more stable but still exhibited substantial shifts. Cross-model agreement dropped from 73% on the bare affirmative framing to 59% under simple negation, indicating that even when models agree on one phrasing, they often diverge when the question is negated.

Human coders who reviewed a sample of responses confirmed that the instability was genuine, not an artifact of measurement. However, they also found that binary agree/disagree proxies overstate the magnitude of the problem, suggesting that an AI model cannot reliably replace human judgment because it silently collapses abstentions and mirrors the forced-choice bias the researchers were studying.

Why Do AI Models Struggle With Negation in Ethical Contexts?

Prior research has established that negation failures are widespread in AI comprehension tasks. Studies have documented systematic insensitivity to negation operators across GPT-style models, with sizable performance drops under diverse negation patterns. One recent benchmark found that models often fail to distinguish true negation from surface-similar distractors. Additionally, research on prompt sensitivity has shown that LLM outputs shift with formatting changes, instruction wording, and example selection, with framing effects persisting from human cognition into AI systems.

What makes this study novel is that it extends these findings beyond semantic comprehension into ethically consequential judgments. Previous work focused on whether models could correctly answer factual questions about negation. This research asks a harder question: when models render moral judgments that people and institutions rely on, does their stance remain stable across logically equivalent phrasings? The answer is no, and the consequences are significant.

Steps to Assess AI Model Stability in Ethical Decision-Making

• Use Polarity-Paired Testing: When evaluating an AI model for use in advisory or decision-making roles, test its responses to the same ethical dilemma phrased both affirmatively and negatively to detect framing sensitivity.
• Apply the Negation Sensitivity Index: Rather than relying on binary agree/disagree measurements, use the NSI metric to directly measure whether a model's ethical stance remains consistent across logically equivalent framings.
• Avoid Single-Phrasing Audits: Do not assess a model's ethical reliability based on a single phrasing of a question, as this can misreport the model's true stance and mask polarity-dependence.
• Supplement with Human Review: Recognize that AI models cannot replace human judgment in high-stakes ethical scenarios, and use human coders to verify that apparent stability is genuine rather than an artifact of measurement.
• Test Across Model Types: Evaluate both commercial and open-weight models, as they exhibit different levels of framing instability, and smaller models may be particularly vulnerable to stance reversals.

What Are the Practical Implications for AI Deployment?

The researchers emphasize a critical limitation: a model whose stance flips with phrasing cannot be relied upon in any high-stakes decision scenario. This does not necessarily mean AI systems should be excluded from advisory roles, but it does mean that institutions deploying these systems must understand their limitations and implement safeguards.

For ordinary users seeking moral advice from chatbots, the research suggests caution. The same question phrased two different ways may yield opposite guidance, each delivered with apparent confidence. For institutions using AI in deliberative roles, the instability means that an analyst's wording choices silently determine the recommendation a committee sees, introducing an unintended source of bias. For researchers and companies using AI models as judges of other models' outputs, the findings reveal a fundamental problem: the measurement instrument itself is biased in ways that can distort benchmark claims about model "values" or "moral alignment".

The research does not address whether this instability extends to operational instructions, such as deployment-time prohibitions in system prompts or policy documents. That remains an open question for future work. However, the evidence presented here suggests that standard robustness evaluations would not catch this problem, because they rarely test polarity-paired framings head to head.