Ottawa is the latest city to turn to AI to predict chronic homelessness

The researcher behind the Ottawa project, Majid Komeili, said the system uses personal data such as age, gender, Indigenous status, citizenship status and whether a person has registered family.

It also looks at factors such as how many times service may have been refused at a shelter and the reasons service was received.

The system will also use external data, such as weather information and economic indicators like the consumer price index and unemployment rate. Komeili said the system will predict how many nights an individual will spend in a shelter in six months.

“This will be a tool in the service providers’ toolbox, ensuring that no one falls through the cracks because of (a) human error. The ultimate decision maker will remain a human,” he said in an email.

This information is available in the first place because people are already “highly screened” for various benefits or treatments, argued McGill University associate professor Renee Sieber.

“Unfortunately, homeless people are extremely heavily monitored, and the data is very intrusive,” Sieber said.

Data may include details about medical visits, addiction, relapses, and HIV status.

Sieber said it’s important to ask whether AI technology is really necessary. “Do you know more about chronic homelessness with AI than you did with a spreadsheet?”

It was only a matter of time before AI got involved, suggested Tim Richter, president of the Canadian Alliance to End Homelessness.

While not widely available, such tools “can probably, to some extent, anticipate who is most likely to experience homelessness or chronic homelessness,” he said. “Using AI to do that could be very useful in targeting interventions to people.”

Most places don’t have good enough data to set up such systems, Richter said.

His organization is working with cities across the country, including London and Ottawa, to help collect better “real-time, person-specific” information — “in a way that protects their privacy.”

Chronic homelessness means that an individual has been homeless for more than six months or has experienced repeated episodes of homelessness over that period.

While 85 percent of people move in and out of homelessness quickly, about 15 to 20 percent “get stuck,” Richter said.

AI systems should be able to do their job and flag individuals who are at risk by analyzing aggregated community-level data and without knowing the specific identity of the individual involved, Richter said.

This is the approach the Ottawa project is taking. Identifiable information such as names and contact information is replaced with codes.

“There is a master list that includes the links between the identifier codes and user identities. AI training and testing operates only on the encoded dataset. The master list is stored separately on a secure server with restricted access,” Komeili explained.

He noted that the system uses data that has already been collected in previous years and is not being collected specifically for use by the AI.

Vinh Nguyen, manager of social policy, research and analysis for the City of Ottawa, said in a statement that any sharing of data collected by the city “undergoes rigorous internal review and scrutiny.”

“The data we share is often aggregated, and when this is not possible, all identifiable information is removed to ensure strict anonymity of users,” he said, adding that collaborations with academics must be reviewed by an ethics board before work with data takes place.

Nguyen said the city is currently conducting “internal testing and validation” and plans to consult with the shelter sector and clients before implementing the model, with consultations planned for late fall.

Alina Turner, co-founder of HelpSeeker, a company that uses AI in products that address social issues, said AI’s “superpowers” ​​could be useful when it comes to comprehensive analysis of the factors and trends that fuel homelessness.

But her company made a conscious choice to stay away from predicting risks at the individual level, she said.

“You can have a lot of bias issues with that,” she said, noting that the data varies across different communities and “the racial bias in that data is also a big challenge.”

A long-recognized problem with AI is that its analysis is only as good as the data it’s fed. This means that when the data comes from a society with systemic racism built into its systems, AI predictions can perpetuate it.

For example, due to systemic factors, indigenous individuals are at greater risk of homelessness.

However, if an AI system automatically gave someone a higher score when they entered a shelter and identified as Indigenous, “there would be a lot of ethical issues with taking that approach,” Turner argued.

Komeili, the Ottawa researcher, said bias is a “known problem with similar AI-powered products.” He noted that humans have biases too, and different individuals may make different recommendations.

“One advantage of an AI-based approach is that when used as an assistive tool in the toolbox of human experts, it can help them converge on a standard approach. Such an assistive tool helps human experts avoid missing important details and can reduce the likelihood of human error.”

Luke Stark, an assistant professor at Western University, is working on a project that studies the use of data and AI for homelessness policies in Canada, including the existing AI initiative in London, Ontario.

He said another issue human decision-makers need to think about is how predictions can cause certain segments of the homeless population to be overlooked.

Women are more likely to avoid shelters for safety reasons and are more likely to turn to options like couch surfing, he noted.

An AI system that uses data from the shelter system will focus on “the kind of people who already use the shelter system… and that leaves out a lot of people.”

Stark described predictive systems as the latest technology that risks obscuring the root causes of homelessness.

“One concern we have is that all this attention to these triage-based solutions takes the pressure off policymakers to actually look at the structural causes of homelessness that are there in the first place,” he said.

As Richter said, “Ultimately, the key to ending homelessness is housing.”

This report by The Canadian Press was first published Aug. 4, 2024.

Anja Karadeglija, Canadian Press