Montreal – part 2: Can urban AI systems help address local needs of vulnerable groups at the neighborhood-scale?

Participation

Understanding how AI systems can contribute to improve the quality of life for people living in cities is key to people-centered approaches to AI and digital technologies.

In the first web story, we shared the methodology and the lessons learnt from the discovery stage of Urban AI’s pilot project in Montreal’s Côte-des-Neiges neighborhood, – led by the Université de Montréal in partnership with the Canadian non-profit Open North – which aimed to understand the local and priority needs of vulnerable and marginalized groups in Côte-des-Neiges and evaluate the potential of urban AI systems as fit-for-purpose solutions. 

In this second chapter, we explore further the results of the neighborhood needs-assessment conducted in 2023 by Open North and the associated opportunities and risks with urban AI systems.

The Urban AI seeks to develop AI projects that are truly adapted to the social and economic realities of local communities in Montreal. The project demonstrates how the AI risk framework presented in the UN-Habitat’s white paper AI & Cities (2022) can be applied in practice, and provides a valuable example of a people-centered approach to AI and community engagement.

A mentoring session to present the pilot project was held on April 24 2024. More information and the recording link are available here.

Using ‘personas’ to illustrate the local needs of vulnerable communities

The first web story described how Open North’s neighborhood needs-assessment used a participatory mixed methodology that cross-references quantitative and qualitative data sources in order to better understand the local needs of vulnerable and marginalized groups in the Côte-des-Neiges neighborhood, which also stands out as one of Montreal’s most populous and multicultural neighborhoods. 

While not an exhaustive list, the study identified nine needs, which were prioritized in the following order of importance by the project’s Steering Committee:

  1. Increasing the supply of affordable and adequate housing;
  2. Reducing language barriers;
  3. Reducing digital divides;
  4. Informing tenants of their rights and obligations;
  5. Improving the discoverability of local services;
  6. Reducing the reliance of low-income households on food aid;
  7. Improving the frequency of public transport through data collection;
  8. Making pedestrian routes, cycle paths and curbs safer;
  9. Reducing the reliance of community organisations fighting food insecurity on food banks.

N.B.: The full neighbourhood study produced by Open North for the University of Montreal provides more details about these needs, the context in which they exist, the potential causes and undesirable effects of the problems that underpin them, and previous attempts to resolve them. This web story only summarizes key points of the study.

To illustrate these needs, Open North developed personas. These fictional archetypes representative of a group’s needs are based on real data, studies, and testimonials from local stakeholders with knowledge of the neighborhood’s reality.

Here are a few examples of these personas:

Trisha, a mother originally from the Philippines, lives in Côte-des-Neiges. She has limited skills in French, but is fluent in English. She feels excluded and isolated, as she cannot understand posters, public announcements or administrative documents available only in French. This limits her ability to obtain information about community events, family support programs or health services available in the neighborhood.

Paul, aged 76, has resided in a 3 1⁄2 bedroom apartment for over 30 years. With a retirement pension of approximately $860 monthly, $550 is allocated for housing. He loves his neighbors and his neighborhood and has no plans to move. However, his landlord informs him of the need to reclaim the apartment for renovations and to house his son. Under persistent pressure, Paul reluctantly signs an agreement barring his return after renovations. Uncomfortable with online apartment searches and limited internet usage, Paul, facing drastically increased rental costs, finds himself without a home on July 1.

Roberto, an employment counselor at a community organization, is dedicated to assisting individuals in finding employment. His clientele often requires support with housing or accessing food aid. Understanding the importance of addressing essential needs for successful employment outcomes, Roberto seeks to collaborate with other neighborhood organizations to ensure holistic support for his members, but lacks actionable information to inform his decision-making. He is convinced that a centralized database or platform listing all available services would enable efficient collaboration, referrals, and ultimately help him in better supporting his members.

Noteworthy considerations from the needs analysis

Three main considerations emerged from the needs analysis in Côte-des-Neiges, which we detail below:

  1. Some needs are considered essential while others are a means to an end
  2. Needs are not evenly distributed in space
  3. Needs are interconnected, requiring a comprehensive approach to problem solving

Distinguishing Essential Needs from Means to an End

As highlighted through the personas, needs often intersect and are not isolated. This emphasizes the necessity of adopting a comprehensive approach to problem-solving.

For example, a significant challenge is that many neighborhood residents in need of local support are unaware of the extensive network of local community organizations offering social services to assist individuals and families meet essential needs. Information dissemination for many groups in the neighborhood relies on informal networks, primarily through word of mouth, and newcomers not part of these networks struggle to access details on support services.

Solutions aimed at enhancing the discoverability of local services do currently exist, such as an online directory of community and social services in Montreal, allowing users to filter their search by category of organization, location, and even service language. While this is a major initiative, the directory isn’t accessible to residents in the neighborhood impacted by digital inequalities (without reliable internet access or with low digital literacy levels), nor is the information actionable to those who don’t speak French (the directory is only provided in French, similarly to most municipal services and administrative forms in Quebec).

In essence, solutions that aim to improve the discoverability of local services must also tackle the digital divide and language barriers in order to be effective.

Needs are not evenly distributed in space

Needs and their underlying problems are not evenly spatially distributed across the neighborhood. We illustrate this in the two maps below. On one hand, map 2 shows that housing unaffordability is more concentrated in the northern portion of the southern neighborhood district known as “Haut de la Côte,” where universities are clustered and adjacent to the expansive Mont-Royal park. On the other hand, map 3illustrates that reliance on food aid in the neighborhood is more concentrated in the northern district referred to as “Bas de la Côte”, according to data from the neighborhood’s primary organization combating food insecurity.

Map 2: Data source: Statistics Canada, Census of population 2021; coordinate and projection systems: WSG84 / UTM N18; Author: Samuel Kohn, Open North.
Map 3: Data source: Multicaf ; coordinate and projection systems : WSG84 / UTM N18 ; Author: Samuel Kohn, Open North

Interconnectedness of needs

The Urban AI pilot project aims to assess the suitability of existing or emerging urban AI systems in addressing identified needs rather than imposing solutions. Embracing a non-techno solutionist approach involves exploring a broad spectrum of solutions, both technological and non-technological, to find the most fitting ones.

When potential urban AI systems are identified, it is important to thoroughly examine their limits. Furthermore, taking inspiration from UN-Habitat’s white paper on AI and cities, a risk-assessment approach should be adopted to evaluate the perceived level of risk associated with the deployment of an AI system within the local context.

In summary, for each identified need, the project sought answers to the following three questions:

  1. Do urban AI solutions effectively address the core need or problem? 
  2. What are the limitations of these solutions, and are there more effective alternatives?
  3. Is the level of risk associated with the AI solutions deemed acceptable within the local context?

Example where urban AI systems do not address the core need and other alternatives are more appropriate: increasing the supply of affordable housing

In the housing sector, AI systems exist for a variety of applications, such as to predict real estate and rent prices to inform housing policy or to predict the risk of chronic homelessness to proactively provide public services based on the level of risk.

While these applications can be useful for specific use cases, none of them directly address the heart of the top need identified: how to actually increase the supply of affordable housing units in the neighborhood? We consider that AI systems have a very limited potential to directly address this need in the context of this study, and that the most appropriate and impactful solutions lie in the political, legal and socio-economic realms, focusing on housing policy and alternative housing models.

Example where urban AI systems effectively address the problem but the level of risk is unacceptable: housing maintenance 

The lack of adequate housing was identified as a top concern in the neighborhood. According to the 2021 census, nearly 10% of dwellings require major repairs, many of which are plagued by unsanitary conditions, such as high humidity levels and the presence of mold. These conditions often lead to health problems, especially among children.

AI systems exist for predictive home maintenance. These systems are designed to sense and monitor the home environment, detect humidity levels, mold, and other issues. While these systems are likely an effective solution to proactively improve housing maintenance, they are extremely invasive and pose significant threats to fundamental rights such as the right to privacy.

Furthermore, these systems would likely be owned and controlled by landowners, and their outputs could be used by the latter to justify “renovictions”*, which are increasingly frequent in the neighborhood and are contributing to the housing crisis.

*”Renovictions” refer to the practice of landlords evicting tenants under the guise of renovations, often to raise rents, which can lead to gentrification and displacement of low-income residents. We consider that these AI systems pose too significant of a risk to vulnerable tenants in the neighborhood, despite its contribution to addressing the problem. 

Can specialized large language models (LLMs) trained on local datasets support local community needs? 

Large language models have become all the hype recently with the advent of tools like Chat GPT. But they are rarely referenced in the literature on urban AI. Given that some of the key needs identified revolve around barriers to information and communication, we explored whether LLMs could be added to the arsenal of fit-for-purpose urban AI systems.

Imagine a virtual AI assistant that is trained on local datasets and the most commonly spoken local languages in order to address local use cases and problems. Using a modular and an iterative approach, use cases such as obtaining information on tenant rights, finding information on local community services, translating administrative forms only available in French, or obtaining information on municipal services could be designed, tested and deployed. This proof of concept is illustrated in the image below, and we believe it has a strong potential to address several needs identified, but also adapt over time to the neighborhood’s ever evolving communication and information needs.

Can urban AI systems help in addressing these needs?

The Urban AI pilot project aims to assess the suitability of existing or emerging urban AI systems in addressing identified needs rather than imposing solutions. Embracing a non-techno solutionist approach involves exploring a broad spectrum of solutions, both technological and non-technological, to find the most fitting ones.

When potential urban AI systems are identified, it is important to thoroughly examine their limits. Furthermore, taking inspiration from UN-Habitat’s white paper on AI and cities, a risk-assessment approach should be adopted to evaluate the perceived level of risk associated with the deployment of an AI system within the local context.

In summary, for each identified need, the project sought answers to the following three questions:

  1. Do urban AI solutions effectively address the core need or problem?
  2. What are the limitations of these solutions, and are there more effective alternatives?
  3. Is the level of risk associated with the AI solutions deemed acceptable within the local context?

 

Example where urban AI systems do not address the core need and other alternatives are more appropriate: increasing the supply of affordable housing

In the housing sector, AI systems exist for a variety of applications, such as to predict real estate and rent prices to inform housing policy or to predict the risk of chronic homelessness to proactively provide public services based on the level of risk.

While these applications can be useful for specific use cases, none of them directly address the heart of the top need identified: how to actually increase the supply of affordable housing units in the neighborhood? We consider that AI systems have a very limited potential to directly address this need in the context of this study, and that the most appropriate and impactful solutions lie in the political, legal and socio-economic realms, focusing on housing policy and alternative housing models.

Example where urban AI systems effectively address the problem but the level of risk is unacceptable: housing maintenance

The lack of adequate housing was identified as a top concern in the neighborhood. According to the 2021 census, nearly 10% of dwellings require major repairs, many of which are plagued by unsanitary conditions, such as high humidity levels and the presence of mold. These conditions often lead to health problems, especially among children.

AI systems exist for predictive home maintenance. These systems are designed to sense and monitor the home environment, detect humidity levels, mold, and other issues. While these systems are likely an effective solution to proactively improve housing maintenance, they are extremely invasive and pose significant threats to fundamental rights such as the right to privacy.

Furthermore, these systems would likely be owned and controlled by landowners, and their outputs could be used by the latter to justify “renovictions”*, which are increasingly frequent in the neighborhood and are contributing to the housing crisis.

*”Renovictions” refer to the practice of landlords evicting tenants under the guise of renovations, often to raise rents, which can lead to gentrification and displacement of low-income residents. We consider that these AI systems pose too significant of a risk to vulnerable tenants in the neighborhood, despite its contribution to addressing the problem. 

Can specialized large language models (LLMs) trained on local datasets support local community needs?

Large language models have become all the hype recently with the advent of tools like Chat GPT. But they are rarely referenced in the literature on urban AI. Given that some of the key needs identified revolve around barriers to information and communication, we explored whether LLMs could be added to the arsenal of fit-for-purpose urban AI systems.

Imagine a virtual AI assistant that is trained on local datasets and the most commonly spoken local languages in order to address local use cases and problems. Using a modular and an iterative approach, use cases such as obtaining information on tenant rights, finding information on local community services, translating administrative forms only available in French, or obtaining information on municipal services could be designed, tested and deployed. This proof of concept is illustrated in the image below, and we believe it has a strong potential to address several needs identified, but also adapt over time to the neighborhood’s ever evolving communication and information needs.

.

Proof of concept, neighborhood virtual assistant (Credits: Samuel Kohn, Jérémy Diaz, Open North)

While we are optimistic about the potential of this proof of concept to address key needs with an acceptable and manageable level of risk, it’s essential to acknowledge the existence of risks and challenges. Deploying such a system doesn’t come with a guarantee of success. The development of Large Language Models (LLMs) itself entails numerous risks and challenges, as outlined in the diagram below. Additionally, persistent issues like the digital divide further complicate matters.

In the next, and final, chapter of this web stories series, recommendations and opportunities are provided to ensure the responsible and ethical design of the AI tool envisioned in the project’s action plan: a virtual neighbourhood assistant trained on local datasets to solve local use cases.

Overview of LLM challenges according to Kaddour et al (2023)
To ensure the responsible and ethical design and development of this tool for the benefit of its future users, Open North has devised a roadmap. This roadmap outlines practical steps for the ethical, inclusive and responsible implementation of this type of urban AI system, which will be discussed in detail in a forthcoming third and final blog post.
Chemin de la Côte-des-Neiges, one of the busiest streets in the neighbourhood (Credits- Jérémy Diaz, Open North)

Jump to

Register and be part of the community of Cities and Digital Rights

Resources

Practical guides for digital transformation in Latin America and the world.

Proof of concept, neighborhood virtual assistant (Credits- Samuel Kohn, Jérémy Diaz, Open North)
Key recommendations: designing and developing AI systems responsibly and ethically for vulnerable and marginalized groups 
jackie-hutchinson-JJYzJXbwB20-unsplash
Montreal - part 1: Exploring urban AI solutions for and with vulnerable communities: a needs-centred approach
Waste Collectors Map
Citizen’s Engagement in Urban Data Initiatives: A typology outlining citizens' engagement in local data initiatives

Beta:

This is a new platform. Leave your comments here to improve it.