Opening the black box: helping AI to persuade without bias

该项目的总体目标是确定人工智能如何使用自然语言以透明和无偏见的方式说服人类。人工智能系统正在成为我们日常生活中不可或缺的一部分，从YouTube或Netflix上的简单推荐，到改变生活的决定，如筛选求职者或推荐贷款。然而，这些系统缺乏透明度和客观的公平指标，产生了难以理解的“建议”，这可能会在整个社会中越来越多地助长有偏见的决定。根据性别区别对待个人的有偏见的建议是一个特别值得关注的问题（例如在就业环境中）。检测和消除这种偏见是极其困难的。它源于刻板的人类标签和分类，这些标签和分类存在于通常用于训练人工智能的庞大拼凑数据集中。因为这种偏见是“隐藏的”，所以它对与AI互动的人类的传播相应地很难确定，甚至更难消除。在这篇博士论文中，我们提出并测试了一种新颖的解决方案，它将实验心理学与计算科学相结合，提供了一种检测人工智能的偏见建议何时转化为人类判断的方法。学生将设计他们自己的人工智能，其中可以“注入”已知的性别偏见模式，这些模式源于我们对认知刻板印象的研究。使用先进的自然语言模型，人工智能将在一系列研究中与人类参与者进行简短的“对话”，这些研究将提供关于拒绝或遵从有偏见和无偏见建议的行为的丰富数据来源。使用尖端的机器学习技术，学生将分析这个新的定制数据集，以识别信号说服的语言特征。在这些特征中，学生将寻求隔离那些与接受有偏见的内容特别相关的特征。至关重要的是，这些特征可以用来检测现实应用中隐藏偏见的传播。

The overall aim of the project is to establish how AI can use natural language to persuade humans in transparent and bias-free ways. AI systems are becoming an integral part of our daily lives, from simple recommendations on YouTube or Netflix, to life-changing decisions such as shortlisting job applicants or recommending loans. However, these systems lack transparency and objective indicators of fairness, producing poorly understood 'advice' that could be fuelling biased decisions on a growing scale right across society. Biased advice that treats individuals differently based on their gender is a particular concern (e.g. in employment contexts). Detecting and removing this bias is extremely difficult. It arises from stereotyped human labelling and categorisation residing within huge, cobbled-together data sets typically used to train AI. Because this bias is 'hidden', its transmission to humans interacting with AI is correspondingly difficult to establish, and even more difficult to eliminate. In this Phd, we propose and test a novel solution which combines experimental psychology with computing science to offer a way of detecting when biased advice from AI passes into human judgements. The student will engineer their own AI into which known patterns of gender-bias, deriving from our work on cognitive stereotypes, can be 'injected'. Using advanced natural language models, the AI will have brief 'conversations' with human participants in a series of studies that will provide a rich source of data on acts of rejection or conformity to biased and unbiased advice. Using cutting-edge machine learning techniques, the student will analyse this new bespoke data set to identify language features that signal persuasion. Within these features the student will seek to isolate ones specifically linked to acceptance of biased content. Crucially, these features may be used to detect the transmission of hidden biases in real-life applications.