内容：
随着人工智能技术的发展，尤其是大型语言模型（LLMs）的崛起，这些模型在处理和生成文本方面展现出了前所未有的能力。然而，一项新的研究表明，即使是像GPT-4这样先进的大语言模型，也可能在误导性信息的影响下产生错误或偏见的输出。
这项研究由清华大学、上海交通大学、斯坦福大学和南洋理工大学的研究人员共同完成，并在国际顶级语言学会议ACL 2024上被接收为口头报告。研究的第一作者是清华大学交叉信息研究院的二年级硕士生许融武，指导老师包括美国东北大学的Weiyan Shi助理教授、清华大学邱寒助理教授和徐葳教授。
研究团队构建了一个名为Farm的数据集，包含1500个事实性问题和相关的误导性信息，以此来测试大语言模型在面对虚假信息时的反应。他们发现，在多轮对话中，这些模型在反复接触误导信息后，有时会自信地做出与事实相悖的判断，如认为地球是平的。
研究人员通过三种劝说策略——晓之以理、树立权威和动之以情——来测试模型的信念变化。结果表明，即使是GPT-4这样的先进模型，也有高达20.7%的可能性被虚假信息所影响。
这项研究对提升大语言模型的安全和鲁棒性具有重要意义。随着生成式人工智能技术的发展，虚假信息的生成和传播变得更加容易，这对社会带来了诸多挑战。研究人员正在探索如何提高模型对虚假信息的识别能力和抵抗能力，以确保这些技术能够被安全、可靠地使用。
总的来说，这项研究提醒我们，虽然大语言模型在文本生成方面表现出色，但它们也存在被误导信息影响的潜在风险。未来，如何平衡模型的能力与安全，将是人工智能领域需要不断探索和解决的问题。
英语如下：
News Title: “AI Can Be Fooled Too: Large Models Lose Faith in Flat Earth Theories”
Keywords: AI Deception, Belief Journey, Fake News
News Content:
Title: New Study Reveals Large Language Models Vulnerable to Misleading Information
As artificial intelligence technology advances, particularly with the rise of large language models (LLMs), these models have demonstrated unprecedented abilities in processing and generating text. However, a new study reveals that even advanced LLM systems like GPT-4 may produce incorrect or biased outputs under the influence of misleading information.
The study, conducted by researchers from Tsinghua University, Shanghai Jiao Tong University, Stanford University, and Nanyang Technological University, was accepted as an oral presentation at the prestigious ACL 2024 international conference on computational linguistics. The first author of the study is Xu Lianwu, a second-year master’s student at the Interdisciplinary Research Institute of Tsinghua University, with guidance from Weiyan Shi, an assistant professor at Northeastern University, Qiu Han, an assistant professor at Tsinghua University, and Xu Wei, a professor.
The research team constructed a dataset named Farm, comprising 1,500 factual questions and related misleading information, to test the reaction of large language models to false information. They found that in multi-round conversations, these models sometimes confidently made judgments contrary to the facts after repeatedly encountering misleading information, such as believing the Earth is flat.
Researchers tested the models’ belief changes using three persuasion strategies: appealing to reason, establishing authority, and appealing to emotion. The results showed that even advanced models like GPT-4 have a 20.7% chance of being influenced by fake information.
This study is significant for enhancing the safety and robustness of large language models. As generative artificial intelligence technology develops, the creation and spread of fake information becomes easier, posing numerous challenges to society. Researchers are exploring ways to improve the models’ ability to identify and resist fake information to ensure these technologies can be used safely and reliably.
In summary, this study reminds us that, while large language models excel in text generation, they also have the potential to be influenced by misleading information. In the future, balancing the capabilities and safety of these models will be an ongoing challenge and exploration in the field of artificial intelligence.
【来源】https://www.jiqizhixin.com/articles/2024-08-06-9
Views: 2