城市的AI权利：公共空间的多元协同设计与治理
The urban right to AI: Pluralistic co-design and governance of public space
浏览论文内容
中文总结 AI 辅助
本论文提出公民AI权利与多元对齐方法，以蒙特利尔为实证基础，结合Street Review与LIVS研究，构建市政AI治理框架，解决公共空间AI决策中的多元价值争议问题。
中文摘要 AI 辅助
城市开始运用AI不仅分析公共空间，还界定关于公共空间的证据范畴。本论文探究当分数、地图及生成图像成为市政决策组成部分时会产生何种结果。本文提出，当代城市运行依托两类耦合基础设施：物质城市与塑造城市感知、比较及行动对象的认知算法层。由于公共空间存在争议，仅靠技术性能无法治理该算法层。本论文提出公民AI权利及多元对齐方法，使公共价值差异得以显现，而非被平均为单一目标。方法论上，本论文在规范理论、参与式研究、机器学习及治理设计间切换。实证研究以蒙特利尔为基础，Street Review将参与式研究与计算机视觉结合，探究居民对街道的评价差异，以及如何用约45000张街景图像在城市尺度上绘制这些判断。LIVS（Local Intersectional Visual Spaces，局部交叉视觉空间）将同一问题扩展至生成式AI，其与30个社区组织共同开发，包含13462张图像的37710组成对比较，用于通过直接偏好优化（Direct Preference Optimization）微调和评估Stable Diffusion XL模型。结果显示对齐可获改善，但分歧与中立性依然存在。本文将这些结果视为注释噪声之外的证据，表明部分价值仍具争议。论文结论将这些发现转化为市政实践，涵盖生命周期治理、采购规则、监督、重新委托及追索权。该框架展示城市如何治理AI，而不将多元价值视为测量误差或强制其纳入单一目标。
英文摘要
Cities are beginning to use AI not only to analyze public space, but also to define what counts as evidence about it. This thesis asks what follows when scores, maps, and generated images become part of municipal decision-making. I argue that contemporary urbanism operates through two coupled infrastructures: the material city and an epistemic, algorithmic layer that shapes what cities can perceive, compare, and act upon. Because public space is contested, this algorithmic layer cannot be governed through technical performance alone. The thesis develops a civic Right to AI and a pluralistic approach to alignment in which differences in public values are made visible rather than averaged into a single objective. Methodologically, the thesis moves between normative theory, participatory research, machine learning, and governance design. The empirical work is grounded in Montréal. Street Review combines participatory research with computer vision to examine how residents evaluate streets differently and how those judgments can be mapped at city scale using approximately 45,000 street-view images. LIVS (Local Intersectional Visual Spaces) extends the same problem to generative AI. Developed with 30 community organizations, it contains 37,710 pairwise comparisons across 13,462 images and is used to fine-tune and evaluate a Stable Diffusion XL model with Direct Preference Optimization. The results show that alignment can improve, but disagreement and neutrality persist. I treat these outcomes not as annotation noise, but as evidence that some values remain contested. The thesis concludes by translating these findings into municipal practice through lifecycle governance, procurement rules, oversight, recommissioning, and recourse. The resulting framework shows how cities can govern AI without treating plural values as measurement error or forcing them into a single objective.
