《经济学人》近期发布研究，通过分析ChatGPT、Claude、Gemini和Grok等大语言模型的输出文本，总结了AI生成内容的识别特征。[1]该研究覆盖55,940个句子和120万词汇，对比分析了AI文本与人类写作的差异。[1]研究发现，AI生成的文本倾向使用更多多音节词汇（如"significant"、"increasingly"、"consequences"）和拉丁词缀，同时使用较少的标点符号且句子结构更长。[1]其中，Claude在文本中更频繁地使用破折号，而ChatGPT的破折号使用明显较少。[1]
不同AI模型的写作特征并非固定不变。[1]佛罗里达州立大学的Tommie Juzek指出，"大语言模型从人类反馈中学习，会采纳人们认为印象深刻的表现手法，放弃不被认可的风格。"[1]这一点在ChatGPT的更新中得到体现——该模型在早期版本几乎每句都使用破折号，而在更新后则表示"应该减少使用"。[1]研究采用Pangram检测工具，其对AI文本的识别准确率声称达99.98%。[1]
The Economist has conducted a comparative analysis of text generated by major language models against human writing, identifying distinctive linguistic patterns that characterize AI-produced content.[1] The research examined 55,940 sentences and 1.2 million words across outputs from ChatGPT, Claude, Gemini, and Grok.[1] According to the study, AI systems consistently favor polysyllabic words such as "significant," "increasingly," and "consequences," alongside an increased use of Latin affixes, fewer punctuation marks, and longer sentence structures.[1] These stylistic markers shift over time as the underlying software evolves.
Notably, different AI models display distinct writing habits. Claude employs em-dashes more frequently, while ChatGPT shows markedly lower usage of this punctuation.[1] ChatGPT's punctuation patterns changed substantially following a system update; the model had previously used em-dashes in nearly every sentence before the revision, after which it adopted more restrained usage.[1] Tommie Juzek from Florida State University explains this shift, noting that "LLMs learn from human feedback, picking up things people find impressive and abandoning things they don't approve of."[1] A detection tool called Pangram claims to identify AI-generated text with 99.98 percent accuracy.[1]