The urban right to AI: Pluralistic co-design and governance of public space

AI systems increasingly mediate how cities perceive, evaluate, and transform public space. When these systems are treated as neutral tools, the algorithmic layer that shapes what becomes legible and actionable in planning is left to market incentives, vendor defaults, or technical convenience. This dissertation argues that contemporary urbanism must be governed as a dual infrastructure: the material city and an epistemic, algorithmic layer that produces representations, scores, and generative alternatives that increasingly influence municipal decisions.

The dissertation develops a civic Right to AI for public-space contexts and operationalizes it through pluralistic alignment. Rather than imposing a universal standard for what counts as an inclusive place, pluralistic alignment requires cities to elicit heterogeneous values, represent them without collapsing them into a single target, and negotiate legitimate standards that can be encoded into both physical design and the AI systems that increasingly mediate urban decisions.

The research addresses five questions concerning: (1) how municipalities should conceptualize the algorithmic layer shaping public space and the responsibilities that follow from it; (2) what a civic Right to AI requires and prohibits in public space, notably with respect to participation, transparency, and avenues for recourse; (3) the extent to which perceptions of streets diverge across social positions and which dimensions are more or less likely to converge through deliberation; (4) how cities can elicit and represent heterogeneous values without collapsing them into a single score, and how such representations change what becomes governable; and (5) how co-produced lifecycle governance and procurement practices can encode negotiated standards into models and streets over time, including mechanisms for recommissioning, audit, and redress.

Empirically, the dissertation draws on participatory studies in Montréal that combine interviews, focus groups, and structured ratings and rankings of streetscape imagery. In one study, 12 participants evaluated 20 streets represented through 60 vantage points, comparing perceptions of inclusivity, accessibility, aesthetics, and practicality between post-occupancy residents and pre-occupancy newcomers. In a second study, 28 residents participated in interviews and then in rating and ranking exercises, including small-group discussions. Across studies, agreement is higher for some visually legible dimensions than for social dimensions such as inclusivity. Structured deliberation increases convergence on selected criteria without eliminating residual contestation, indicating that disagreement should be documented and governed rather than treated as annotation noise.

Technically, the dissertation develops Street Review as a participatory measurement infrastructure and demonstrates how locally co-produced labels can support subgroup-aware predictive modeling and citywide mapping. The Street Review pipeline combines participatory elicitation of evaluative descriptors with a scalable model that predicts multi-criteria streetscape evaluations and produces citywide heatmaps from approximately 45,000 street-view images. The results show that scaling with co-produced labels can surface spatial patterns useful for auditing and prioritization, but also that image data quality and representational limits constrain what can be governed through vision-only inputs.

The dissertation also introduces LIVS (Local Intersectional Visual Spaces), a pluralistic alignment dataset for inclusive public spaces developed through a two-year participatory process with 30 community organizations. LIVS encodes 37,710 pairwise comparisons across 13,462 images, structured along six criteria derived from 634 community-defined concepts. Using Direct Preference Optimization (DPO), the work fine-tunes a Stable Diffusion XL model and evaluates the fine-tuned model through case studies that examine preference alignment and neutrality. As part of a resident evaluation workshop, 2,100 additional annotations were collected. Of these, 700 judgments favored the DPO model, 300 favored the baseline, and 1,100 were neutral. These results indicate that preference tuning can improve alignment on a subset of judgments while leaving a large indeterminate region that should be treated as a governance signal rather than forced into optimization.

The dissertation synthesizes these findings into an implementation architecture: a Right to AI as a municipal governance layer, an augmented co-produced AI lifecycle with checkpoints from co-framing to co-maintenance, and procurement and oversight mechanisms that enable auditability, recommissioning, and recourse. Together, these elements specify how cities can govern the algorithmic layer of public space while treating pluralism as an empirical condition and a requirement of legitimacy.

Keywords: urban planning; public space; AI governance; participatory AI; pluralistic alignment; intersectionality; deliberation; preference learning.