Linguistic Ambiguity and Hallucination in Large Language Models: A Cross-Linguistic and Cultural Analysis of Output Accuracy

An inaccurate response to an ambiguous prompt can reflect a mistaken interpretation, an unsupported factual assertion or a translation problem. These outcomes require different explanations and evaluation criteria. This paper examines their relationship through a structured literature analysis and a proposed Chinese-English evaluation of GPT and Claude models. It distinguishes ambiguity in meaning and reference from factual support, translation fidelity and the adequacy of available evidence. Existing studies inform a comparative framework, while their different models, languages and tasks limit direct comparisons. The proposed evaluation pairs ambiguous prompts with disambiguated versions and contrasts implicit cultural context with explicit context. Bilingual reviewers validate the materials before generation. Exact model versions, prompting conditions and output records are retained, and blinded human annotation separates supported claims, confirmed contradictions, unsupported additions and unresolved cases. Evaluation reports factual outcomes alongside answer coverage, clarification and abstention, with repeated outputs analysed at the item level. Mitigation involves clarifying consequential uncertainty, supplying relevant context and checking claims against appropriate sources. The paper contains no new model outputs, measured hallucination rates or model rankings. Its contribution is a design for examining how language and context affect interpretation and factual reporting without treating every incorrect or unscorable response as hallucination.