When Choices Become Risks: Safety Failures of Large Language Models under Multiple-Choice Constraints

2026-04-18

AI Summary

This study addresses a critical safety alignment vulnerability in large language models (LLMs) when deployed in structured tasks such as multiple-choice questions, where the inability to abstain from answering can lead to unsafe responses. The authors systematically evaluate 14 mainstream LLMs using adversarially constructed multiple-choice questions in which all available options are unsafe. Their analysis reveals, for the first time, an inverted U-shaped relationship between constraint strength and violation rate. Furthermore, they demonstrate that test questions generated by high-capability models exhibit strong cross-model transferability and near-saturating violation rates. These findings indicate that forced-choice mechanisms substantially increase the likelihood of policy-violating outputs, highlighting a significant underestimation of potential risks in current safety evaluations for structured interaction scenarios.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness

Application Category

• Security and Privacy: Large-scale security measurements
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Safety alignment in large language models (LLMs) is primarily evaluated under open-ended generation, where models can mitigate risk by refusing to respond. In contrast, many real-world applications place LLMs in structured decision-making tasks, such as multiple-choice questions (MCQs), where abstention is discouraged or unavailable. We identify a systematic failure mode in this setting: reformulating harmful requests as forced-choice MCQs, where all options are unsafe, can systematically bypass refusal behavior, even in models that consistently reject equivalent open-ended prompts. Across 14 proprietary and open-source models, we show that forced-choice constraints sharply increase policy-violating responses. Notably, for human-authored MCQs, violation rates follow an inverted U-shaped trend with respect to structural constraint strength, peaking under intermediate task specifications, whereas MCQs generated by high-capability models yield near-saturation violation rates across constraints and exhibit strong cross-model transferability. Our findings reveal that current safety evaluations substantially underestimate risks in structured task settings and highlight constrained decision-making as a critical and underexplored surface for alignment failures.

Problem

• safety alignment
• large language models
• multiple-choice constraints
• structured decision-making
• refusal behavior

Innovation

• safety alignment
• multiple-choice constraints
• forced-choice failure
• structured decision-making
• policy violation