康奈尔大学研究人员发现，主流AI聊天机器人在生成创意故事时存在一个奇特现象：它们频繁创造并重复使用同一虚构人物伊莱亚斯·索恩（Elias Thorne）1。研究团队分析了约20000个来自OpenAI、Anthropic和Google等公司的大语言模型生成的故事，发现包含"伊莱亚斯"、"玛拉"、"埃拉拉"等特定名字以及"灯塔看守人"、"钟表匠"、"图书管理员"等职业的故事占比高达88%1。其中，伊莱亚斯作为灯塔看守人的设定出现在近三分之二的故事中1。
这一虚构人物的影响范围已扩展到互联网其他领域。伊莱亚斯·索恩的形象出现在Amazon上的AI图书、音乐列表、YouTube视频和健康指南中1。研究人员认为这种现象源于两个方面：AI安全对齐训练减少了模型可用的数据来源池，以及大语言模型之间的交叉培训导致信息重复1。
Researchers at Cornell University have uncovered a peculiar phenomenon in which mainstream artificial intelligence chatbots consistently create and reuse an identical fictional character named Elias Thorne across thousands of generated stories.1 An analysis of approximately 20,000 narratives produced by large language models from OpenAI, Anthropic, and Google revealed that specific names like "Elias," alongside professions such as "lighthouse keeper," "watchmaker," and "librarian," appeared in 88 percent of the stories examined.1 Most strikingly, the character of Elias working as a lighthouse keeper surfaced in roughly two-thirds of the analyzed narratives.1
The fictitious figure has transcended AI chatbot conversations and infiltrated other digital platforms, with Elias Thorne appearing in AI-generated books available on Amazon, music playlists, YouTube videos, and health guides.1 Cornell researchers attribute this repetitive pattern to two interconnected factors: safety alignment training in AI systems has constrained the pool of training data available to these models, and cross-training between different large language models has amplified the duplication of information across systems.1
评论
还没有评论，欢迎留下第一条。