营销公司Graphite发布研究报告，通过分析前沿AI模型的文本生成特征，揭示了这些系统在写作中留下的明显痕迹1。该研究基于10000篇ChatGPT发布前的人类文章作为对照组，由AI模型从摘要重写以消除源文本偏差1。
研究发现，尽管各实验室已在改进明显特征，但AI模型仍保留着广泛且难以完全消除的语言特征1。具体而言，Graphite识别出13000个短语在AI内容中的使用频率至少是人类内容的两倍1。Claude Opus 5.5表现最为显著：单词"dependable"的出现频率比人类样本高23倍，短语"this matters"高116倍，"why X matters"高92倍1。OpenAI Astra则过度使用"another dimension"和"may provide"/"can provide"等表述，其纠正框架构造出现频率超过100倍1。
各模型在改进em-dash使用方面已取得进展1。其中，Opus 5.5相比Opus 5减少了99%的em-dash使用，Astra减少了88%，而Gemini 3.1 Pro几乎完全消除了这一特征1。然而，这些改进仍未能根除AI文本的整体识别特征，表明完全消除模型文本指纹面临持续的技术挑战。
Marketing firm Graphite has released research identifying persistent linguistic signatures in cutting-edge artificial intelligence models, revealing that despite industry efforts to minimize such tells, AI-generated text continues to exhibit recognizable stylistic quirks.1
The study examined writing patterns across leading models including Claude Opus 5.5 and OpenAI Astra, discovering that 13,000 phrases appear in AI content at frequencies at least double those found in human writing.1 Claude Opus 5.5 demonstrates particularly pronounced tendencies, with the word "dependable" appearing 23 times more frequently than in human samples, while phrases like "this matters" occur 116 times more often and "why X matters" 92 times more often.1 OpenAI Astra exhibits its own distinctive markers, including excessive use of "another dimension" and constructions like "may provide" or "can provide," with correction-framing patterns appearing over 100 times more frequently than in human text.1
The research methodology relied on 10,000 pre-ChatGPT human articles as a baseline, with AI models rewriting from summaries to eliminate source text bias.1 While AI developers have made measurable progress in reducing some obvious indicators—Opus 5.5 decreased em-dash usage by 99 percent compared to its predecessor Opus 5, Astra reduced such usage by 88 percent, and Gemini 3.1 Pro nearly eliminated the pattern entirely—the research suggests that broader stylistic fingerprints remain difficult to fully erase.1
评论
还没有评论，欢迎留下第一条。