Department of Life Sciences Communication, University of Wisconsin-Madison, Madison, USA.
Department of Computer Science, Stanford University, Stanford, USA.
Sci Rep. 2024 Jan 18;14(1):1561. doi: 10.1038/s41598-024-51969-w.
Autoregressive language models, which use deep learning to produce human-like texts, have surged in prevalence. Despite advances in these models, concerns arise about their equity across diverse populations. While AI fairness is discussed widely, metrics to measure equity in dialogue systems are lacking. This paper presents a framework, rooted in deliberative democracy and science communication studies, to evaluate equity in human-AI communication. Using it, we conducted an algorithm auditing study to examine how GPT-3 responded to different populations who vary in sociodemographic backgrounds and viewpoints on crucial science and social issues: climate change and the Black Lives Matter (BLM) movement. We analyzed 20,000 dialogues with 3290 participants differing in gender, race, education, and opinions. We found a substantively worse user experience among the opinion minority groups (e.g., climate deniers, racists) and the education minority groups; however, these groups changed attitudes toward supporting BLM and climate change efforts much more compared to other social groups after the chat. GPT-3 used more negative expressions when responding to the education and opinion minority groups. We discuss the social-technological implications of our findings for a conversational AI system that centralizes diversity, equity, and inclusion.
自回归语言模型利用深度学习生成类人文本，其应用日益广泛。尽管这些模型取得了进步，但人们对它们在不同人群中的公平性仍存在担忧。尽管人工智能公平性已被广泛讨论，但对话系统中的公平性衡量标准仍存在不足。本文提出了一个框架，该框架植根于审议式民主和科学传播研究，用于评估人机通信中的公平性。我们使用该框架进行了算法审核研究，以检查 GPT-3 如何回应在关键科学和社会问题（如气候变化和“黑人的命也是命”运动）上存在社会人口背景和观点差异的不同人群：气候变化和“黑人的命也是命”（BLM）运动。我们分析了 3290 名参与者与 20000 次对话，这些参与者在性别、种族、教育程度和观点上存在差异。我们发现，在意见少数群体（如气候变化否认者、种族主义者）和教育少数群体中，用户体验明显较差；然而，与其他社会群体相比，这些群体在聊天后对支持 BLM 和气候变化努力的态度发生了更大的变化。GPT-3 在回应教育和意见少数群体时使用了更多的负面表达。我们讨论了我们的发现对一个重视多样性、公平性和包容性的对话式人工智能系统的社会技术影响。
