Recently, I travelled to Italy to join a delegation tasked with exploring how AI might be used to support ending homelessness. The convening was called by Community Solutions, a US based non-profit that pioneered the Built for Zero methodology, and hosted by the Rockefeller Foundation.

Through a facilitated process we explored the opportunities and limits of AI in housing systems, drawing on expertise from global perspectives across research, government, nonprofits and tech. 

The time I spent with this interdisciplinary group, away from my usual routine and environment and its associated competing priorities and distractions, inspired and clarified new ideas across AI, homelessness and their intersection:

1. Explainability in LLMs (Large Language Models)

It is commonly accepted that it is not possible to explain how LLMs produce their outputs. This makes it difficult to assess ethical concerns, such as bias, discrimination and verification of source input data. As Mary Gray, Senior Principal Researcher at Microsoft and Faculty Associate at Harvard University’s Berkman Klein Centre for Internet and Society explained, it doesn’t have to be this way! Lack of explainability is an engineering choice that has been followed due to a perceived lack of demand for enhanced explainability. This might include visualisation tools, feature attribution methods, and counterfactual explanations, which help show the set of data and probabilities used to generate outputs. .As buyers of technology we must push our vendors for AI systems that are more transparents.

2. Assembling multidisciplinary teams can develop ideas

Bringing together experts from various fields, such as computer science, social work, urban planning and policy allows for ideas to be progressed rapidly. Diverse expertise can for example help sense check technical assumptions or culture and social considerations. Getting the right people in the room for ideation is a powerful tool and something which we should draw on more in London local government.

3. Operate at both the macro and micro level to end homelessness

As parts of the world become increasingly uninhabitable due to climate change, the convening highlighted the growing crisis of climate-induced migration, a new factor that will likely exacerbate homelessness. It’s this type of macro level trend for which we need data combined with data on rents, house prices and house building that is needed to help us design long term housing policy. At the same time we need to respond to factors that are affecting people’s lives now that can lead to homelessness. These can include life events such as evictions, job loss or interaction with specific public services . Through data aggregation, research and analysis these data sets may converge to provide insights that support both perspectives of the agenda. 

4. Certain types of predictive modelling are unlikely to work in the context of homelessness due to the limits of the data we have

While some recent projects have identified data sources which seem to indicate certain indicators are accurate predictors of some forms of future homelessness we need to be aware of what’s missing in the data we have on people who experience homelessness. When using Machine Learning techniques have been applied to areas such as predicting which children may be taken into care the models were judged to be ineffective. Modelling complex social dynamics is very difficult, not because of the technology or analytical approach but because the data required to capture the nuances of human behaviour is vast and complicated by how quickly human circumstances change. We need to consider what is realistic, ethical and technically viable to collect when we approach any predictive work in this space. 

5. It is possible to obtain (some) eviction data

Data on evictions through the court system has proved difficult and time-consuming to obtain as county court data is not published centrally. In the US, nonprofit New America tackled this problem through the Foreclosure and Evictions and Analysis Tool that enables practitioners, campaigners and the public to analyse  eviction data from distributed data sources. It’s a model with insights and lessons that should inform a similar approach in the UK that could build a new data source to inform our understanding of homelessness and rough sleeping.

What happens next

As part of the convening the group developed a number of ideas that are now being considered for further funding. One idea that resonated with me builds on work already taking place in a number of London Borough social care teams. It uses AI to take notes from natural language conversations between front line or outreach workers and service users. The AI transcription and summarisation helps caseworkers to have more natural conversations with service users while capturing the structured data required to complete current forms and processes.

Beyond improving the interaction and experience of case work and service users and improving data quality it may open up a new data set. Within the raw text of the unstructured conversations there will be many nuances likely missed in the structured data that is usually collected to fulfil many administrative requirements. Using ML featuring extraction we may be able to build a new data set that includes insights into housing and homelessness that are not currently captured in existing data systems. 

Research will be required to evaluate this data but the hope is that the insights derived may help us to redesign housing systems and their accompanying systems to make them better able to serve residents and communities.