💻 computer science

Pluralistic-Alignment Urbanism: Operationalizing a Right to AI for Inclusive Public Space

This paper proposes Pluralistic-Alignment Urbanism (PAU), a procedural governance framework that operationalizes a Right to AI for municipal public-space systems by leveraging participatory case studies in Montreal to transform structured disagreement and subgroup variations into concrete governance mechanisms like disaggregated reporting, versioned value registers, and deliberative oversight.

Original authors: Rashid Mushkani

Published 2026-06-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Rashid Mushkani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a city as a giant, living puzzle. For a long time, city planners have tried to solve this puzzle by looking at the pieces and assigning them a single score: "Is this street safe? Is it pretty? Is it accessible?" They used to do this with clipboards and surveys. Now, they are using Artificial Intelligence (AI) to do it faster, scanning thousands of street photos to generate maps, scores, and even fake pictures of what a street could look like after a renovation.

The problem, as this paper points out, is that everyone sees the puzzle differently. A street might feel "safe" to one person but "scary" to another. A "pretty" park might feel "exclusionary" to someone else. When AI tries to give every street one single number or one single "best" picture, it accidentally erases these different viewpoints. It pretends there is only one right answer, when in reality, there are many.

This paper proposes a new way to run the city's AI, called Pluralistic-Alignment Urbanism (PAU). Think of it as a new rulebook for how cities should use AI, based on the idea that disagreement is not a mistake to be fixed, but a fact of life to be managed.

Here is how the paper breaks it down, using simple analogies:

1. The Problem: The "One-Size-Fits-All" Trap

Currently, cities use AI like a judge with a single gavel. The AI looks at a street, applies a set of rules, and bangs the gavel: "This street gets a score of 7 out of 10."

The Issue: This score tries to combine "safety," "beauty," and "accessibility" into one number. But these things don't always go together. A street might be very beautiful but hard to walk on for someone in a wheelchair. If you average them out, you lose the specific truth about who feels excluded.
The Paper's Claim: You cannot treat public space like a math problem with one correct answer. The "truth" depends on who is looking.

2. The Solution: A "Group Chat" Instead of a Gavel

The authors suggest treating AI systems not as judges, but as facilitators of a massive, ongoing town hall meeting. They call this a "Procedural Right to AI."

Instead of asking the AI to pick the "best" answer, the system is designed to:

Listen to different groups: It asks different people (based on their backgrounds) how they see the street.
Keep the disagreement visible: Instead of averaging the scores into one number, it keeps the different scores separate. It says, "Group A thinks this is safe; Group B thinks it is not."
Respect the "I don't know" answer: When people look at a generated picture of a new park and say, "I can't tell which one is better," the system treats that "neutral" answer as important data, not as a glitch to be ignored.

3. The Two Experiments (The "Proof of Concept")

The authors tested this idea in Montreal, Canada, using two different tools:

Tool A: "Street Review" (The Scorecard)

What it did: They asked residents to describe what makes a street inclusive. They turned these descriptions into a checklist (Accessibility, Aesthetics, Practicality, Inclusivity). Then, they trained an AI to look at street photos and predict how different groups would rate them.
The Result: The AI got really good at predicting the scores (it was 89% accurate). But the key finding was that the scores were different for different groups.
The Lesson: The AI proved that you can scale up these different viewpoints to map the whole city, but you must keep the maps "disaggregated." You can't just show one map; you need to show the map for "Accessibility" and the map for "Safety" separately, because they don't always match.

Tool B: "LIVS" (The Image Generator)

What it did: They used AI to generate pictures of what a street could look like. They asked people to pick the "better" picture based on specific criteria.
The Result: When people were asked to choose, over 50% of the time, they said "Neutral" or "I can't decide."
The Lesson: This wasn't a failure of the AI. It was a signal! It meant that for half the comparisons, the pictures were too similar, or the criteria were too vague to pick a winner. The paper argues that cities should treat this high "neutral" rate as a stop sign. It means, "Don't use this AI to make a final decision here; we need more human discussion."

4. The New Rulebook: How Cities Should Actually Use AI

The paper proposes a governance system (PAU) that acts like a safety harness for these AI tools. Here are the main rules:

The "Value Register": Imagine a living document that lists what the city cares about. It clearly separates things that are easy to measure (like "how many trees are there?") from things that are contested (like "does this feel welcoming?"). The AI is only allowed to make decisions on the easy stuff. For the contested stuff, it just reports the different opinions.
The "Deliberative Cell": This is a standing group of regular citizens who meet regularly to review the AI's work. If the AI starts making weird predictions or if people are confused, this group has the power to say, "Pause. Stop using this tool until we figure it out."
The "Pause and Rollback" Button: If the AI starts causing harm or if the data shows it's lying about what a street feels like, the city has the authority to immediately turn it off and go back to the previous version.
No "Black Box" Decisions: The AI cannot be used to make individual decisions (like denying a permit to a specific person) or to enforce laws. It is only for planning and brainstorming.

The Big Takeaway

The paper concludes that AI in cities shouldn't try to solve the argument about what a "good" city looks like. That argument is healthy and necessary.

Instead, AI should be a tool that keeps the argument visible. It should show us where people disagree, why they disagree, and when the computer is just guessing. By treating disagreement as a feature rather than a bug, cities can use AI to build spaces that are truly inclusive, rather than spaces that just look good on a spreadsheet.

In short: Don't ask the AI to be the mayor. Ask the AI to be the scribe that writes down everyone's different opinions, so the real people can make the final decision.

Technical Summary: Pluralistic-Alignment Urbanism (PAU)

Problem Statement

Municipal agencies increasingly deploy machine learning systems to inventory public spaces, score streetscapes, and generate synthetic imagery for urban planning. These systems function as an "epistemic and algorithmic layer" that converts heterogeneous human experiences into comparable artifacts (scores, maps, images) to inform budgeting and design. However, public space is characterized by persistent disagreement regarding inclusion, safety, and belonging. Current governance frameworks often treat these systems as neutral decision-support tools, relying on single evaluation targets that assume a universal consensus. This approach risks obscuring contested values, erasing subgroup differences through majority averaging, and legitimizing decisions that fail to account for the pluralistic nature of public life. The paper argues that treating disagreement as "noise" to be eliminated is a governance failure; instead, disagreement and indeterminacy must be treated as decision-relevant evidence.

Methodology

The paper proposes Pluralistic-Alignment Urbanism (PAU), a procedural governance framework operationalized through two participatory case studies conducted in collaboration with community organizations in Montréal, Canada. The methodology treats participation not as post-hoc validation but as the co-production of evaluation criteria, data, and governance artifacts.

Street Review (Representational System):
- Goal: To evaluate streetscapes based on resident criteria and scale these judgments citywide.
- Process:
  - Elicitation: Semi-directed interviews with 28 residents generated 600+ descriptors, consolidated into four criteria: Accessibility, Aesthetics, Practicality, and Inclusivity.
  - Evaluation: 12 participants rated 60 street-view images individually, then discussed them in facilitated groups to produce collective judgments.
  - Scaling: A supervised multi-output regression model was trained on 120 rated images (grouped by street-level data points to prevent leakage) to predict scores for 45,000 street images. The model predicts 28 scores (4 criteria × 6 self-identified subgroups + 4 aggregates).
- Metrics: Predictive performance measured via $R^2$ on held-out test sets.
LIVS (Local Intersectional Visual Spaces) (Generative System):
- Goal: To align text-to-image models with pluralistic preferences for public-space design.
- Process:
  - Data Construction: Workshops with community partners elicited 634 concepts, consolidated into six criteria (Accessibility, Safety, Comfort, Invitingness, Inclusivity, Diversity). 440 participant-authored prompts were used to generate 16,693 images via Stable Diffusion XL.
  - Preference Collection: 18 participants annotated 37,710 image-pair comparisons using a web-based platform. Crucially, the interface included an explicit neutral option for indifference or indeterminacy.
  - Alignment: A Direct Preference Optimization (DPO) fine-tune was performed on 35,510 non-neutral comparisons.
  - Evaluation: A held-out workshop with 25 residents provided 2,100 new comparisons to evaluate the tuned model against the baseline, specifically tracking the rate of neutral selections.

Key Contributions

Pluralistic-Alignment Urbanism (PAU): A governance framework that treats public-space AI systems as civic infrastructure requiring a procedural Right to AI. This right is not a general access right but a set of minimum procedures (entitlements, duties, artifacts, recourse) for systems that shape collective conditions.
The Evidentiary Object: A conceptual shift viewing the coupled artifact (stimulus, criteria, protocol, aggregation, representation) as the unit of governance. Disagreement and neutrality are properties of this object, not errors.
Governance Architecture: A concrete implementation model featuring:
- Disaggregated Reporting: Avoiding default majority averaging; reporting subgroup-specific evidence.
- Versioned Value Register: A municipal record distinguishing benchmarkable dimensions from contested ones.
- Standing Deliberative Cells: Participatory bodies with the authority to update the value register and recommend pauses or rollbacks.
- Pause and Rollback Authority: Explicit triggers for halting systems based on documented harms, drift, or indeterminacy.

Results

The empirical studies yielded four propositions connecting technical constraints to governance implications:

P1: Pluralism is Structured, Not Noise.
- Criteria are not commensurable. Correlations between criteria were moderate and uneven (e.g., Practicality and Aesthetics were uncorrelated at $r = -0.05$ ; Inclusivity correlated moderately with Accessibility and Aesthetics).
- Subgroup ratings on the same stimuli differed significantly, indicating that aggregation choices are consequential and can erase accountability-relevant differences.
P2: Deliberation Changes the Evidentiary Object.
- Group judgments were not simple averages of individual ratings. Discussion led to reinterpretations of criteria and new reasons for scoring.
- Residual disagreement after deliberation was recorded as evidence constraining permissible model uses, rather than being forced into consensus.
P3: Co-produced Scaling is Feasible but Bounded.
- The Street Review model achieved $R^2 = 0.91$ (validation) and $R^2 = 0.89$ (held-out test).
- However, the modality (street-view imagery) limits representation to visible cues, failing to capture experiential factors like harassment or symbolic belonging. The resulting maps are audit artifacts, not ground truth.
P4: Neutrality in Preference Alignment is a Governance Signal.
- In the LIVS held-out evaluation (2,100 comparisons), 52.4% of selections were neutral (95% CI [50.2, 54.5]).
- DPO-tuned outputs were preferred in only 33.3% of cases.
- High neutrality rates indicate that many comparisons do not support a determinate ordering under the given criteria, serving as a signal to constrain optimization rather than a label to be smoothed.

Significance and Claims

The paper claims that the legitimacy of public-space AI systems depends less on predictive performance alone and more on whether institutions can sustain ongoing revision, contestation, and scope control.

Operationalizing a Right to AI: The paper translates abstract rights into enforceable municipal procedures, including procurement clauses that mandate disaggregated reporting and pause authority.
Reframing Indeterminacy: By treating neutrality and disagreement as inputs to governance rather than failures of the model, PAU provides a mechanism to justify what systems can and cannot legitimately represent.
Institutional Robustness: The proposed architecture (deliberative cells, value registers) aims to prevent "governance capture" by dominant norms and ensures that AI-mediated evidence remains contestable.

The authors explicitly state that the goal is not to resolve disagreement about public space, but to govern systems in ways that keep such disagreement visible, contestable, and consequential. The framework is presented as a procedural floor for municipal decision-making, acknowledging that the specific criteria and value registers must be locally re-commissioned and cannot be transferred as fixed templates.