CAREER: The Impact of Associations and Biases in Generative AI on Society

该项目旨在巩固生成性人工智能(AI)的伦理基础。产生式人工智能系统建立在语言、语音和视觉机器学习模型的单模和多模组合之上。生成性人工智能模型，如聊天机器人ChatGPT和文本到图像生成器稳定扩散，提供了创新和实用的工具。然而，这项技术存在固有的问题。生成性人工智能模型从大规模的社会文化数据中学习认知心理学中记录的隐含联系和偏见，这是人类关于性别、种族或民族、社会阶层、年龄、能力、性取向、国籍、宗教、概念和跨部门联系的偏见的来源。生成性人工智能偏见对人工智能的绩效差异以及人工智能伦理，特别是关于生成性人工智能对个人和社会的影响产生了影响。来自易于访问的生成性AI模型的输出包含有偏见的社会关联，放大了复杂的偏见，对于AI开发人员和用户来说，缓解这些偏见都具有挑战性。该项目将开发评估生成性人工智能中的关联和偏见的方法，评估生成性人工智能对社会的影响，并分析生成性人工智能如何塑造人类认知和代理。该项目将通过开发方法来解决机器、人类-人工智能合作和社会中的偏见，从而促进民权。该奖项颁发的开源工具和材料将提高一系列利益相关者的认识，包括不同的学生群体、研究人员、开发人员、行业、开源社区、人工智能用户、政策制定者和公众。这一努力将通过引入跨学科和部门的生成性人工智能伦理课程来加强华盛顿大学的人工智能教育。随着人工智能监管和立法的制定，该奖项产生的科学证据将为政策制定者提供安全、可靠和可信的人工智能开发和使用方面的信息。与政策智库的积极合作将有助于将知识转移到决策中。该项目将把计算机和信息科学在机器学习、自然语言处理、计算机视觉、语音处理和人类-人工智能交互方面的研究与社会认知的方法论和大规模数据集结合起来。该项目的主要目标是对生成性人工智能的社会影响进行实证分析，为人工智能的伦理和负责任的开发和部署做出贡献。该项目寻求通过开发原则性和可推广的检测和测量方法来评估和表征生成性人工智能系统中的关联和偏见。利用这些发现和开发的偏差评估方法，研究将设计出自动识别和减少生成性人工智能模型中的偏差信号的方法，并考虑到特定的任务、应用、上下文和用例。这些方法将包括诸如训练数据扩充、嵌入空间处理、微调、指令调整和从反馈中强化学习等技术。通过在州和国家层面比较人类的隐性和显性人工智能偏见，并确定新出现的生产性人工智能偏见，该项目将揭示生殖性人工智能的更广泛的社会影响。分析人类内隐联想测试分数和决策在暴露于有偏见的生成性人工智能输出后的变化，将评估它们对人类在人类-人工智能协作中的偏见、感知和决策的影响。因此，该奖项将开发新的方法，通过引入新的协会来减轻生成性人工智能偏见造成的负面后果，使生成性人工智能与人类价值保持一致。这一奖项带来的潜在进步超越了计算机和信息科学，为认知科学、心理学、语言学、社会学和政治学提供了工具和见解，同时为哲学、法律和政策等领域提供了信息。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project aims to solidify the foundations of ethics in generative artificial intelligence (AI). Generative AI systems are built on unimodal and multimodal combinations of language, speech, and vision machine learning models. Generative AI models such as the chatbot ChatGPT and the text-to-image generator Stable Diffusion offer innovative and practical tools. However, this technology has inherent problems. Generative AI models learn implicit associations and biases documented in cognitive psychology from large-scale sociocultural data, which is a source of human biases regarding gender, race or ethnicity, social class, age, ability, sexuality, nationality, religion, concepts, and intersectional associations. Generative AI bias poses implications for performance disparities in AI, as well as AI ethics, particularly concerning the impact of generative AI on individuals and society. Outputs from easily accessible generative AI models contain biased social associations, amplifying complex biases that are challenging to mitigate for both AI developers and users. This project will develop methods for evaluating associations and biases in generative AI, assess the impact of generative AI on society, and analyze how generative AI shapes human cognition and agency. The project will advance civil rights by developing methods to address bias in machines, human-AI collaboration, and society. The open-source tools and materials presented by this award will raise awareness among a range of stakeholders, including the diverse student population, researchers, developers, industry, the open-source community, AI users, policymakers, and the public. This effort will enhance AI education at the University of Washington by introducing a generative AI ethics curriculum across disciplines and divisions. As AI regulation and legislation are being formulated, the scientific evidence produced by this award will inform policymakers on the safe, secure, and trustworthy development and use of AI. Active collaboration with policy think-tanks will aid transfer of the knowledge to policymaking.This project will integrate computer and information science research in machine learning, natural language processing, computer vision, speech processing, and human-AI interaction with methodologies and large-scale datasets from social cognition. The project's primary objective is to empirically analyze the societal impact of generative AI, contributing to the ethical and responsible development and deployment of AI. The project seeks to evaluate and characterize associations and biases in generative AI systems by developing principled and generalizable detection and measurement methods. Leveraging the findings and developed bias evaluation methods, the research will devise approaches that automatically identify and reduce bias signals in generative AI models, taking into account the specific task, application, context, and use case. These approaches will encompass techniques such as training data augmentation, embedding space processing, fine-tuning, instruction tuning, and reinforcement learning from feedback. By examining generative AI biases in comparison to implicit and explicit biases of humans at the state and country levels and identifying emergent generative AI biases, the project will uncover the broader societal impact of generative AI. The analysis of changes in human implicit association test scores and decisions following exposure to biased generative AI outputs will assess their influence on human biases, perception, and decisions in human-AI collaboration. Accordingly, the award will develop novel approaches that align generative AI with human values by introducing new associations to mitigate the negative consequences caused by generative AI biases. The potential advancements resulting from this award extend beyond computer and information science, providing tools and insights for cognitive science, psychology, linguistics, sociology, and political science, while informing fields such as philosophy, law, and policy.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.