一项学术研究调查了人工智能生成的文本为何相对难以理解的原因1。研究团队分析了九个前沿和近前沿的AI模型，发现这些模型相比人类写作以及回译生成的文本，更频繁使用目的分词、长句法依存、低频n-gram等心理语言学难度特征1。具体来看，所有研究模型相比人类写作都使用了更多无生命主语搭配有生命动词和低频n-gram，而大多数模型还采用了更长的句法依存、堆叠名词和减少的目的分词1。
研究团队提出了一项核心假设来解释这一现象：人类语言针对成功沟通进行了优化，而AI文本则针对数学、代码和推理等目标进行优化，这导致了对信息密度而非沟通易懂性的压力1。研究人员通过对语言模型施加长度惩罚的微调方法验证了这一假设1，进一步确认了AI生成文本与人类表达方式之间存在的根本性差异。
Academic researchers have identified specific linguistic patterns that make text generated by artificial intelligence more difficult to comprehend compared to human writing.1 An investigation into nine leading and near-leading AI models revealed that these systems consistently employ psycholinguistic features associated with greater complexity, including inanimate subjects paired with animate verbs, longer syntactic dependencies, stacked noun phrases, and low-frequency n-grams.1 By contrast, human-written text and passages generated through back-translation showed notably fewer instances of these challenging linguistic structures.1
The research team proposed that this divergence stems from fundamentally different optimization objectives between human and artificial language.1 While human language has evolved to prioritize successful communication and comprehension, AI text appears optimized for mathematical reasoning, code generation, and abstract inference—objectives that place pressure on information density rather than communicative clarity.1 Researchers verified this hypothesis through fine-tuning adjustments that imposed length penalties on language models, with findings that corroborated the proposed mechanism.1
评论
还没有评论，欢迎留下第一条。