【作者】 周瑾宜； 黄英辉； 李伟卿； 董王昊； 邓莎莎； 王伟军；

【Author】 Jinyi Zhou;Yinghui Huang;Wanghao Dong;Weiqing Li;Shasha Deng;Weijun Wang;School of Psychology, Central China Normal University;School of Management, Wuhan University of Technology;School of Economics and Management, Hubei University of Technology;School of International Business Administration, Shanghai International Studies University;

【机构】 华中师范大学心理学院； 武汉理工大学管理学院； 湖北工业大学经济管理学院； 上海外国语大学国际工商管理学院；

【摘要】 随着生成式人工智能(AI)技术的重大突破,其产生的在线虚假评论的数量可能会大幅度增长,进而对以电子口碑为核心的网络消费信息内容生态构成潜在威胁。亟需深入分析人类与生成式AI产生的虚假评论在语言行为上的差异,并提出针对性的识别方法及治理策略。本研究以人类评论者及ChatGPT为研究对象,从欺骗识别领域中的认知负荷理论、人际欺骗理论及现实监控理论的整合视角出发,构建认知负荷、确定性、情感、事件的距离、感知情境细节、认知过程六类语言线索体系,利用计算语言学及机器学习方法构建人类真实评论、人类虚假评论与生成式AI产生的虚假评论的解释性分类预测模型,并比较三者对关键性语言线索使用的差别,探讨生成式AI与人类在欺骗策略使用上的异同。实验结果表明,分类预测模型的准确率、精确度、召回率、F值和AUC-ROC值分别达到89.24%、90.77%、90.23%、90.23%和98.05%。认知负荷(0.0361:p<0.001)、感知情境细节(0.0159:p<0.001)、情感(0.0072:p<0.001)等传统欺骗识别理论相关语言线索,以及所发现的非正式语言(0.0097:p<0.001)、个人关切(0.0091:p<0.001)语言线索均能显著提高模型的AUC-ROC值。针对三者在语言线索使用上的比较分析结果发现,生成式AI产生的虚假评论表现出最大的认知负荷(p<0.001)、最多的情绪表达(p<0.001)和个人关切(p<0.001),且更具正式性(p<0.001)与确定性(p<0.01)。在欺骗策略的使用方面,一方面,人类和AI生成的虚假评论均表现出相似的较低水平的认知负荷;另一方面,生成式AI更多地使用了与人类不同的欺骗策略:相对更多的个人关切相关语言线索(p<0.001)、积极和消极情绪表达(p<0.001);相对更少认知过程语言线索(p<0.01)。此外,通过干预认知负荷,能够有效从部分维度区分真实和虚假评论,这种干预对识别AI产生的虚假评论尤其有效。本研究探索了人类与生成式AI在虚假评论言语行为的可区分性及表现差异,验证了欺骗识别相关理论在生成式AI产生的虚假评论识别中应用价值,有利于网络平台管理机构及企业制定更好的虚假评论监管机制,以应对生成式AI对网络消费信息内容生态造成的潜在重大风险。更多还原

【Abstract】 With the breakthrough progress of generative artificial intelligence technology, the online fake reviews it generates may grow massively, potentially seriously threatening the online consumer information content ecology.It is urgent to deeply analyze the differences in language behavior between human and generative artificial intelligence(AI) reviews, and propose targeted fake review dectection methods and governance strategies. This study takes human reviewers and ChatGPT as research objects, based on the perspectives of cognitive load theory,interpersonal deception theory, and reality monitoring theory in the field of deception detection, integrates cognitive load, uncertainty, emotion, distance, perceived contextual details, cognitive process of language clues,using word embedding and machine learning methods to build explanatory classification prediction models of authentic and fake reviews of human, and fake review of generative AI, compare the differences in the use of language clues among the three, explore the similarities and differences in deception strategy use between generative AI and humans. The experimental results show that the accuracy, precision, recall, F-value, and AUC-ROC value of the classification prediction model are 89.24%, 90.77%, 90.23%, 90.23%, and 98.05%respectively. Language cues related to traditional deception detection theories such as cognitive load(0.0361:p<0.001), perceived contextual details(0.0159: p<0.001), and emotions(0.0072: p<0.001), as well as the discovered language cues of informal language(0.0097: p<0.001) and personal concerns(0.0091: p<0.001) can significantly improve the ROC AUC value of the model. The comparative analysis results of language clue usage behavior found that the fake reviews generated by generative AI show the greatest cognitive load(p<0.001), the most emotional expression(p<0.001) and personal concern(p<0.001), and are more formal(p<0.001) and certain(p<0.01). In terms of the use of deception strategies, both human and AI-generated fake reviews show similar low levels of cognitive load. However, generative AI uses more deception strategies different from humans: relatively more personal concern-related language clues(p<0.001), positive and negative emotional expressions(p<0.001);relatively fewer cognitive process language clues(p<0.01). In addition, by intervening with cognitive load, it is possible to effectively differentiate between real and fake reviews in some of the dimensions, and this intervention is particularly effective in identifying AI-generated fake reviews. This study explores the distinguishability and performance differences of fake review verbal behavior between humans and generative AI, verifies the value of deception detection related theories in the dectection of fake reviews generated by AI, and is beneficial for online platform management agencies and enterprises to formulate better supervision mechanisms to cope with the potential major risks of generative artificial intelligence to the network information ecology.更多还原