Reliability of Large Language Models Depends on Question Constraints

Large language models (LLMs) are frequently evaluated based on visible failure cases such as hallucinations or incorrect outputs. This paper argues that such evaluations are often incomplete, as they overlook the role of question structure and constraint in shaping model behavior. By analyzing context compression, underconstrained queries, and interaction design, the paper demonstrates that many perceived failures arise from ill-posed questions rather than intrinsic model unreliability. It introduces the concept of an inverse survivor bias, in which unsuccessful interactions are disproportionately observed while successful AI usage remains largely invisible. The paper proposes a constraint-based framework for interpreting AI outputs and emphasizes that improving question design and user understanding is a more effective and lower-cost approach than enforcing correctness through increasingly restrictive systems.