Street Scenes: Public Appliances for GenAI Video in Informal Settlements

Paper Title

Street Scenes: Public Appliances for GenAI Video in Informal Settlements

Publication Info

• Topic area: Generative AI applications in low-resource, informal urban settlements.
• Keywords: Generative AI, informal settlements, public appliances, self-expression, community engagement, accessibility, Dharavi, video generation, participatory design, governance.

Background and Problem

• Problem / challenge: Generative AI tools for creativity and self-expression are largely inaccessible to residents of informal settlements due to high costs, literacy barriers, and infrastructure limitations. Existing AI systems are not designed for low-resource, public, or shared contexts.
• Significance: Addressing this gap enables equitable access to AI technologies, fostering creativity, self-expression, and community engagement in underserved areas.
• Motivation and related work: Previous studies like "Hole in the Wall" and StreetWise explored public access to emerging technologies in informal settlements. However, these focused on information access rather than creative authorship. Current generative AI research is predominantly Global North-centric, with limited exploration of its use in resource-constrained environments.

Solution

• Proposed approach: Street Scenes, a public appliance enabling walk-up interaction with generative AI to create short videos using images, voice, and physical prompts.
• Novelty:
  1. Introduction of a public appliance model for generative AI, distinct from private device-based systems.
  2. Empirical insights into how informal settlement residents engage with generative AI for play, identity, business promotion, and community messaging.
  3. Design lessons for accessible, equitable, and community-governed AI systems.
• Procedure and key techniques:
  • Conducted ideation workshops in Mumbai and Nairobi to gather user perspectives.
  • Developed and iteratively refined two Wizard-of-Oz prototypes to test usability and interaction modes.
  • Deployed a fully functional prototype in Dharavi for 13 days, enabling public use in two locations (a photocopy shop and a juice kiosk).
  • Integrated multimodal inputs (images, voice, buttons, dials) and used commercial AI services for video generation.
  • Employed safeguards like content moderation and multilingual support to address local constraints.

Results

• Concrete findings:
  • 779 sessions initiated during deployment; 194 videos successfully generated (69.1% in "Bring to Life" mode, 30.9% in "Inspire Me" mode).
  • Average video rating: 3.44/5; most common uploaded image type: individual portraits (67.4%).
  • Common video purposes: self-presentation (11.3%), creative re-contextualization (16.5%), advertising (10.8%), and community assistance (2.1%).
• Advantage over baselines:
  • Enabled access to generative AI for users with limited literacy and technological experience.
  • Fostered social engagement and collective learning through public gallery features and group interactions.
  • Addressed infrastructural challenges by replacing Bluetooth with direct camera-based image capture.
• Experiments / evaluation:
  • Two rounds of Wizard-of-Oz prototyping with 30 participants (15 per round) to refine design.
  • In-situ deployment in Dharavi with 13 days of operation across two public locations.
  • Mixed-method analysis combining interaction logs, observational data, and qualitative feedback.
• Limitations and future work:
  • Limited to Dharavi, with findings not fully generalizable to other contexts.
  • Short deployment duration (13 days) limits understanding of long-term use and governance.
  • Speech recognition struggled with local dialects; future work should address multilingual and dialect-sensitive systems.
  • Ethical concerns around privacy, misuse, and moderation require deeper exploration.

Summary

This paper introduces Street Scenes, a public appliance for generative AI video creation, designed to overcome barriers of cost, literacy, and infrastructure in informal settlements like Dharavi, Mumbai. Through workshops, prototyping, and a 13-day deployment, the study demonstrates how residents used the system for self-expression, community messaging, and small business promotion. Findings highlight the potential of public AI appliances as shared media platforms, but also reveal challenges around privacy, governance, and accessibility. The study contributes a model for accessible AI design, empirical insights into community engagement, and actionable lessons for embedding governance and sustainability in public AI systems.