Beyond the Townhall

Carina I Hausladen, Javier Argota Sánchez-Vaquerizo, Michael Siebenmann,
Arthur Capozzi, Sachit Mahajan, Dirk Helbing

An Integrated Information Environment for 

Participatory Urban Planning

In democracies around the world,
citizens increasingly decide upon sustainability projects.

• Switzerland
  • CHF 4.9 bn motorway expansion (2024)
• Paris
  • scooters banned (2023)
  • SUV parking tripled (2024)
  • 500 streets pedestrianised (2025)

Experiment

Spatial anchoring

Dual coding

Event Segmentation

• source-grounded information provision
   
• Roles
  • Simple information provision
  • Facilitate reflective dialogue

Results

Results

n = 195

Results

Treatment

Control

Consultation quality changed

Recall

Results

• 500 residents contributed their ideas through surveys.
• 35 businesses added their perspectives.
• 40 focus group members joined two design workshops
• All of this input was compiled into Idea Notebook.

LLM Perception

Results

Conversations

Results

2 of 1,578 | 0.13%

Out-of-fact-pack

• "Fruit trees can turn the area into a meeting place, a place for strangers to share a fruit, start a conversation."
   
• "I think some car spaces could be reserved for disabled people who need them."
   
• "I am concerned about the funding of this project and how other budgets may be affected by this"

Visual Empathy

• Strong | Names the beneficiary (people, children, environment) and imagines future benefit
  • "Fruit trees can turn the area into a meeting place, a place for strangers to share a fruit, start a conversation."
  • "Foster a stronger community and help people to feel a sense of belonging."
• Weak | General benefit, no named beneficiary / 3rd person
  • "It will be a much more pleasant area to visit."
  • "Make the area seem more calm and welcoming and a nicer place to be and hang out."

Voting

Results

Entrenched Cases

Results

Entrenched Cases

Limitations

Limitations

• Participants were paid online workers (Prolific) reasoning about a city in which they had no real stake.
• Sample cohort was predominantly White (85%), UK-based, educated

Conclusion

• Test how information provision (video vs text) changes LLM interaction

• Treatment improves recall

• LLMs were rated as helpful and trustworthy

• Mixed opinions on whether roles should be split or not

• Participants like interacting with the LLMs

• Entrenched cases did not move

• LLMs were reliable

• Treatment turned reflective dialogue more often towards empathy

• tool for cities to test info package before going live
• make interaction with LLMs optional
• give citizens the choice of interaction with different personas

Implications

What's next?

Wasserbus Konstanz

• Citizen deliberation scheduled for October 
• Common practice: Send one page of information; have one day of in person discussion; then recommendation
•  

• How can you better integrate information and 1:1 tutoring?
• How does the "democratic ideal" look like?

Spatial anchoring

Dual coding

Event Segmentation

• Can you make the "immersion" cheaper?
• Can you build one tool that automates this process for the shared information baseline? 
• e.g. do you have a drone? maybe a few short videos enhanced with video generation / image augmentation are enough?

Appendix

StudyTopicNew?

Let me know if you'd like this as a downloadable Word or Excel file instead.

• Public trust in governmental AI remains low
• Normative legitimacy question
  • Democratic communication is not judged solely by factual correctness
  • Foucault (1980) emphasize that discourse is never fully neutral, but is embedded within historically structured power relations
    • moto-normativity: presenting "both sides" of an issue does not guarantee neutrality
• ​Design question
  • accurate information <-> reflective dialogue
• Can not be solved theoretically; they require context-specific and democratically grounded judgment

Source-grounded

• Project facts
• Voting process
• Project cost
• Construction timeline
• Public transport
• Accessibility

• 65 comments from 36 users
• Sentiment coding:
  • 44% negative  
  • 34% neutral
  • 22% positive

• Wasteful public spending
• Spending benefits only a tiny minority
• Cyclist "lawlessness" and pedestrian safety
• cyclists should pay their way

• Voting patterns did not change
• Entrenched positions: Five participants remained immovably hostile.
  • conversational AI is not a substitute for face-to-face deliberation in entrenched cases. 

Limitations

• RAG = facts live in a large external store; a retriever searches per question and injects only the matching chunks
• Our approach = the entire one-page fact pack is pasted into the prompt, in full, every turn, for every participant
  • No search step, no index, no retriever — the "corpus" is small enough to hand over whole

Source-grounded vs RAG

• You are Flo Normale, the official chatbot of Ville d'Ordinaire.
• Your role is to provide factual information about the urban renovation project, grounded only in the provided fact pack.
• If a question matches the fact pack, answer directly and clearly.
• If the requested fact is missing, share any related facts you do know, ask whether that information is satisfying, and emphasize that the question is important and should also be asked to Gustavo.
• If the participant asks for interpretation or opinion, explain that you provide facts only and suggest asking Gustavo, who is designed to help with interpretation and sense-making. If the question is unclear, ask a clarifying question.
• Flo should follow a principle of progressive disclosure. Initial responses should always be short, concrete, and factual, consisting of no more than one or two sentences. If the participant asks for clarification or more information, provide additional detail, context, and background. Avoid unnecessary evasion, and do not say "I don't have details" when approximate or general factual information is available.
• Maintain a neutral, calm, professional, and slightly warm tone. Use simple, accessible language throughout. At the end of each interaction, encourage further questions and invite the participant to consider whether they have enough factual information to make an informed vote, for example by asking, "Would you like more detail, or is this enough for now?"

You are Gustavo, the dialogue partner of Ville d'Ordinaire. Your role is to listen carefully to participants, reflect back their impressions, and encourage them to think through their opinions about the proposed city planning elements before they vote in the city-wide referendum.
Engage participants in an active and thoughtful conversation. Encourage them to explore their views by asking open-ended follow-up questions that help clarify their thinking. Always reflect back what participants have said to demonstrate understanding. Remain neutral at all times: do not judge their views, steer them toward a particular choice, or suggest that one opinion is better than another. Instead, help participants consider trade-offs from multiple perspectives.
On first mention of a topic, keep responses brief, using no more than two or three sentences. If participants ask follow-up questions or request more depth, provide a more detailed response, expanding into multiple short paragraphs that further develop the discussion.
Whenever possible, ground reflections in the provided fact pack. If a question falls outside the available facts, remain engaged by drawing on any related information that may be relevant. Rather than presenting facts as definitive answers, use them as prompts for reflection. For example, instead of simply stating, "The renovation costs CHF 414,000," say, "The renovation is estimated at CHF 414,000. How does that figure strike you in terms of value for the community?"
Encourage participants to think not only about what they believe, but also why they hold those views. Ask reflective, meta-level questions such as, "Why do you think that's important?" or "What makes that aspect stand out to you?" The overall goal is to foster metacognition by helping participants connect factual information with their own values, priorities, and perspective before making their decision in the referendum.

Gustavo

other(s) · people · community/communities · neighbour(s) · child/children/kids · elderly · senior · resident(s) · visitor(s) · tourist(s) · wildlife · social · belonging · together · gather/gathering · generation(s) · everyone · each  other · family/families · stranger(s)

Social/beneficiary lexicon

• Real-world comparison based on an urban planning project in Lausanne:
  • 47 online comments from 36 users
• Sentiment classification: multilingual

  • cardiffnlp/twitter-xlm-roberta-base-sentiment model

• Online comments: 50.7% negative | 28.4% positive | 20.9% neutral
• LLM interactions  : 10.3% negative | 37.0% positive | 52.6% neutral

Comparison to real online discussion

From Self-Interest to Collective Reasoning

Prosocial Urban Planning

Carina I Hausladen

Effects of Immersive Information on

Information

• Public trust in governmental AI remains low
  • largely due to concerns about hallucinations, misinformation, and bias
  • mitigation: Retrieval-Augmented Generation (RAG) and explicit source citation
• Normative legitimacy question
  • Democratic communication is not judged solely by factual correctness
  • Foucault (1980) emphasize that discourse is never fully neutral, but is embedded within historically structured power relations
    • moto-normativity: presenting "both sides" of an issue does not guarantee neutrality
• ​Design question
  • accurate information <-> reflective dialogue

Can not be solved theoretically; they require context-specific and democratically grounded judgment

Coding and Validation

Coding and Validation

Phase 1: Codebook Development

• Authors independently coded source material
• LLM coding as additional input (Nguyen-Trung, 2025)
• Codebook refined iteratively until LLM misclassifications disappeared

Phase 2: Multi-Coder Application & Adjudication

• Three independent coding passes: Claude Opus, Claude Sonnet, OpenAI GABRIEL (Sentence-level coding; treatment/control shuffled)
• Adjudication: disagreements resolved using codebook

Phase 3: Independent Human Validation (in progress)

• Train independent coders on codebook
• Compare human codes to LLM codes

Human-LLM Qualitative Coding

Conversations changed

Treatment

Control

Old Introduction

• Technical innovation alone is not sufficient for urban sustainability transformations.
• Effective solutions require meaningful public participation.

Participatory Urban Planning

Participatory Urban Planning

Participatory Urban Planning

Information

• Participation cannot be meaningful without understanding
• Well-designed briefing materials have been shown to reduce polarization and enable reflective judgment (Fishkin 2018)
• Information must be engaging, and memorable.

Information

engaging

Information

memorable

• Meaningful democratic participation requires that citizens recall information to deliberate and vote. 
• The design of information material must align with the
  cognitive mechanisms of memory

Information

Literature

Literature

 Human-LLM-interaction

• Studies condition on generally-known, attitude-laden topics: abortion, partisanship, conspiracies (Salvi et al. 2025; Costello et al.  2024; Hackenburg et al. 2024, 2025; Bai et al. 2025; Argyle et al. 2025; Goldstein et al. 2024; Durmus et al. 2024).
• We hold causal control over the information participants bring in.