💻 computer science

Collective Recourse for Generative Urban Visualizations

This paper proposes "collective recourse," a structured framework for communities to submit visual bug reports that trigger targeted fixes in generative urban visualization models, demonstrating that while prompt-level adjustments offer speed, deeper interventions like dataset edits and reward-model tweaks provide more durable solutions with higher planner adoption.

Original authors: Rashid Mushkani

Published 2026-05-04

📖 5 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 planner using a high-tech "magic paintbrush" (an AI image generator) to show citizens what a new park or street might look like. The problem is, this magic paintbrush sometimes makes mistakes. It might accidentally leave out certain groups of people, show stereotypes, or depict neighborhoods in a way that feels unfair or unsafe.

This paper proposes a new system called Collective Recourse. Think of it as a "Visual Bug Report" system for the city, but instead of one person complaining, it's the whole community working together to fix the paintbrush.

Here is how the system works, broken down into simple parts:

1. The Problem: The Magic Paintbrush Gets Biased

Just like a real paintbrush might have a broken bristle that always leaves a smudge, the AI has "glitches." If you ask it to draw a busy street, it might keep drawing only one type of person or miss important local details. These aren't just small errors; they can hurt feelings, spread stereotypes, and influence how people vote on city plans.

2. The Solution: A Community Repair Shop

Instead of waiting for the AI company to notice the problem, the city sets up a repair pipeline where citizens act as the quality control team.

The Report: When a group of people sees a bad image, they don't just say, "That's wrong." They file a structured report. They say: "Here is the picture, here is who is missing or being misrepresented, and here is how many of us saw this."
The Scorecard: The system gives every report a "Mandate Score." It's like a traffic light.
- Severity: How bad is the mistake?
- Volume: How many people reported it?
- Representativeness: Does this group actually represent the whole neighborhood, or is it just a loud few?
- Evidence: Is there proof?
- If the score is high enough, the city must fix it.

3. The Four Tools to Fix the Paintbrush

The paper tests four different ways to fix the AI, comparing them like different tools in a toolbox:

The "Prompt Patch" (Counter/Negative Prompts): This is like adding a sticky note to the paintbrush instructions. "Make sure to include a wheelchair ramp" or "Do not draw only men."
- Pros: Super fast (done in 2–3 days).
- Cons: The AI often forgets the note later, and the mistake comes back (high recurrence).
The "Data Diet" (Dataset Edits): This is like teaching the paintbrush by showing it new, better pictures. If the AI didn't know what a specific neighborhood looked like, you add those photos to its training library.
- Pros: Very durable. The fix sticks (low recurrence).
- Cons: Takes longer (about 2 weeks).
The "Reward Tuner" (Reward Model Tweaks): This is like training the paintbrush to "feel" good when it gets things right and "feel" bad when it gets them wrong.
- Pros: Also very durable and helps city planners actually use the new rules.
- Cons: Takes the longest (about 3 weeks).

4. The "Citizen Jury"

To make sure the fix is fair, the city doesn't just let the computer decide. They use a Rotating Citizen Jury. Imagine a group of regular neighbors taking turns to review the fixes. They check: "Did we actually fix the problem? Is the new picture fair to everyone?" This prevents the system from just listening to the loudest voices and ensures the fix works for the whole community.

5. What the Tests Showed

The researchers ran a simulation with 240 fake reports to see how this would work in real life.

Speed vs. Quality: Quick fixes (sticky notes) were fast but the problems came back often. Slow fixes (changing the training data) took longer but solved the problem for good.
The Sweet Spot: They found a "Goldilocks" score (0.12) for the Mandate Score. If they set the bar too high, they missed real problems. If they set it too low, they got overwhelmed with false alarms. At this level, they caught 75% of the real problems with very few mistakes.
Fairness: When they made sure the reports came from a diverse mix of people (not just one loud group), they caught even more problems without losing accuracy.

The Big Picture

This paper argues that we shouldn't just treat AI as a black box that we can't touch. Instead, we should treat it like a public service tool. By giving communities a structured way to report errors and forcing the city to fix them based on clear rules, we can make sure the "magic paintbrush" draws a city that looks like everyone lives there, not just a stereotype.

In short: It's a system where the community acts as the editors, the city acts as the publisher, and the AI is the writer that keeps getting corrected until the story is fair for everyone.

1. Problem Statement

Text-to-image diffusion models (e.g., Stable Diffusion XL) are increasingly used in urban planning to visualize streetscapes, infill options, and design guidelines. However, these models suffer from systematic social biases, including demographic stereotyping, occupational gender bias, and the omission of specific communities.

The Gap: Current algorithmic recourse frameworks focus on individual counterfactual explanations (how a single user can change an outcome). This is insufficient for generative urban imagery, where harms are typically group-level, recurring, and structural.
The Challenge: There is no established mechanism for communities to collectively report visual bugs, trigger specific model fixes, and ensure those fixes propagate into municipal planning workflows (e.g., design guides, procurement).

2. Methodology

The authors propose a framework called Collective Recourse, which treats community inputs as structured "visual bug reports" that trigger a formal mitigation pipeline.

A. The Collective Recourse Pipeline

The system operates through a five-stage workflow:

Report: Communities submit structured reports containing the prompt, generated image, context, harm type, severity, and evidence.
Triage: An intake team clusters reports by harm type and context. They calculate a Mandate Score to determine if a fix is required.
Fix Selection: Based on the triage, one of four Recourse Primitives is selected:
- Counter-prompts: Adding protective, context-aware language.
- Negative prompts: Excluding unwanted motifs.
- Dataset edits: Curating/augmenting training data to fix coverage gaps.
- Reward-model tweaks: Updating preference models (similar to RLHF) to steer generation.
Verification: Fixes are tested against a Subgroup Visual Eval Set (SVES)—a set of templated prompts seeded with urban-perception cues—to measure recurrence.
Closure & Governance: Verified fixes are logged on public dashboards, update Model Cards/Datasheets, and are linked to planning artifacts (design guides, RFQs). Rotating citizen juries review closures to ensure representativeness and adjudicate disputes.

B. The Mandate Score ( $M$ )

To decide when a report triggers a mandatory fix, the authors define a score combining severity, volume, representativeness, and evidence quality:
$M(b) = s \cdot (1 - e^{-n/\tau_n}) \cdot r \cdot q$
Where:

$s$ : Severity ($0-1$)
$n$ : Count of unique reporters (Volume saturation)
$r$ : Representativeness index (alignment with population marginals)
$q$ : Evidence quality
$\tau_n$ : Saturation parameter

A mitigation is mandated when $M(b) \geq \tau_M$ (a threshold tuned for precision/recall trade-offs).

C. Evaluation Strategy

The authors conducted a simulation study using a synthetic program of 240 reports across six subgroups (Women, Seniors, LGBTQ+, Newcomers, African diaspora, Neighborhood organizations). They tracked:

Mitigation Latency: Time from intake to verified pass.
Recurrence: Probability of the harm reappearing within 30 days.
Policy Adoption: Whether planners adopted the fix in official documents.

3. Key Contributions

Formalization of Collective Recourse: A new framework shifting from individual to group-level recourse, linking model lifecycle management with municipal governance.
Typology of Recourse Primitives: A classification of four fix types (Counter-prompts, Negative prompts, Dataset edits, Reward tweaks) with distinct speed-durability profiles.
Mandate Thresholds: A mathematical definition for triggering institutional action based on collective signals, balancing urgency (recall) and rigor (precision).
Integrated Pipeline: A practical design connecting technical fixes to participatory planning workflows, including citizen juries and public dashboards.

4. Key Results (Simulation)

The simulation compared the four recourse primitives and tested different mandate thresholds:

Speed vs. Durability Trade-off:
- Prompt-level fixes (Counter/Negative prompts): Fastest (Median 2.1–3.4 days) but least durable, with high recurrence rates (21–38%).
- Structural fixes (Dataset edits/Reward tweaks): Slower (Median 13.5–21.9 days) but significantly more durable, with lower recurrence (12–18%) and higher policy adoption (30–36%).
Threshold Sensitivity:
- A threshold of $\tau_M = 0.12$ yielded 93% precision and 75% recall.
- Increasing representativeness ( $r$ ) via rotating juries improved recall to 81% with negligible precision loss, suggesting procedural innovations can improve equity without sacrificing accuracy.
Subgroup Variance: Latency varied by subgroup (2.8 to 4.7 days), with Seniors showing the lowest recurrence (13.9%) and Women experiencing slightly higher latency (4.7 days), highlighting the need for stratified service levels.

5. Significance and Implications

Bridging Technical and Civic Domains: The paper moves beyond "transparency as an end goal" (e.g., Model Cards) to "transparency as a mechanism for action." It operationalizes accountability by linking model fixes directly to planning artifacts like Request for Qualifications (RFQs) and design guidelines.
Mitigating Group Harms: By focusing on collective signals rather than individual counterfactuals, the framework addresses systemic biases that affect entire communities in urban visualization.
Governance Innovation: The use of rotating citizen juries and mandate thresholds provides a scalable, democratic method for auditing AI in public sectors, preventing "overfitting to vocal groups" while ensuring marginalized voices are heard.
Practical Roadmap: The study offers a concrete "division of labor" for cities: use prompt-level fixes for immediate, temporary relief, but prioritize dataset and reward-model changes for long-term, structural correction.

In conclusion, the paper argues that for generative AI to be safe and equitable in urban planning, it requires a collective recourse infrastructure that transforms community complaints into enforceable, auditable, and durable technical and policy changes.