CHAI (chronic-homelessness prediction)

CHAI, the Chronic Homelessness Artificial Intelligence tool, is a machine-learning model built and run by the City of London, Ontario. About six months ahead, it flags people in the city's shelter system likely to become chronically homeless: more than 180 shelter days in a year. Caseworkers see each flag with the facts about the person that drove it.

How CHAI was built

The City's Municipal Artificial Intelligence Applications Lab, within its Information Technology Services, built CHAI in-house. An external consultant supports maintenance and retrains the model about once a year. The code is published open source under an MIT license, which lets anyone reuse it.

A September 2020 preprint by Blake VanBerlo and colleagues describes the technical work. It sets out two designs of neural network. One is a multilayer perceptron, a standard design. The other combines a person's service history over time with fixed demographic facts. Both were trained on records of 6,521 de-identified people, whose names and identifying details were removed. The records were grouped in 30-day steps.

CHAI went live in August 2020, during the city's response to COVID-19. It cost roughly CA$57,000 to build, and about CA$1,100 a month to run.

What it draws on

CHAI draws on de-identified records from HIFIS, the federal Homeless Individuals and Families Information System. The records are pooled across about 20 to 24 London homelessness-support organizations. In production it was described as drawing on about 21 million data points. Early use identified 88 people as at risk.

Who decides

The caseworker does. City materials framed CHAI as providing intelligence to caseworkers rather than making decisions itself. It makes no service, priority, or eligibility decision.

Consent and explanations

Two design choices are distinctive and genuine. First, CHAI is consent-based. It uses information about people who are in HIFIS, and lets them opt out of having their data included. It was built with reference to the principles of GDPR, the European Union's data-protection law. It was also built with reference to Canada's Directive on Automated Decision-Making, the federal government's rules for automated decision systems.

Second, CHAI was designed to be interpretable. It gives local explanations, which show which of a person's features most affected that person's score. So caseworkers can see what drove a flag. The builders framed this as building trust and reducing unintended bias, instead of operating as a black box.

The accuracy figure

During testing, CHAI was reported to predict chronic homelessness with "93 percent accuracy." The technical work gives two more telling figures, from 10-fold cross-validation on historical data. That test trains the model ten times, each time holding back a tenth of the records to check it against.

Recall, the share of real cases the model found, was about 0.921. Precision, the share of flagged people who were real cases, was about 0.651. So about a third of the people it flagged in testing were not real cases. The case file notes that precision sits well below recall, and few people in the records become chronically homeless. So a single accuracy figure overstates how reliable the flags are.

No independent validation after deployment, of accuracy or of what happened to people flagged, was found.

Why the city built it

The city justified CHAI with cost figures. A chronically homeless person was said to cost roughly CA$135,000 a year. Such a person uses about 534 shelter days, against about 45 for someone who is not chronically homeless. That is roughly twelve times as many.

Who the data leaves out

In 2023 Global News, a Canadian news organization, commissioned an independent review. Two unaffiliated computer-science experts and a privacy lawyer found CHAI appeared to take the steps needed to protect personal information. They also noted it remains "a model of surveillance."

They noted that HIFIS-based data can miss vulnerable groups: aging people, women with children, new immigrants, Indigenous families, and users of private shelter systems. CBC News, reporting at CHAI's launch, raised the same point. A public-shelter definition of homelessness leaves out people who couch-surf or make private arrangements, disproportionately women.

The opt-out tightens this frame. Each opt-out removes a person from the group CHAI can see, and CHAI has no record of the opt-out itself.

Research and wider debate

An ongoing research project at Western University's Starling Centre situates CHAI in its historical, legal, and policy context. It raises concerns about fairness, accountability, surveillance, and inequality.

By 2024, as Ottawa and other cities adopted similar tools, London was described as the Canadian pioneer of this kind of prediction. Coverage raised the concern that raising a person's risk score for identifying as Indigenous is problematic. It also noted that predictive tools can obscure the structural root causes of homelessness and take pressure off policy-makers. Tim Richter, president of the Canadian Alliance to End Homelessness, stressed that the key to ending homelessness is housing.

What this network is drawn from

This network follows the pattern the case file describes. It is not a reconstruction of the actual tool. It shows the partner agencies' records, HIFIS, CHAI, its flags, the caseworker, and the outside reviewers. The people CHAI scores, and the people it cannot see, are outside the network.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Here a mistake might be a person wrongly flagged, or one at real risk left unflagged. Closing a failure pathway means mistakes stop passing along it. The work along it goes on: caseworkers still read flags.

The Lab shows the network in one of three states. Self-correcting means mistakes are caught faster than they spread. Cascading failure means they multiply faster than they are corrected. A tipping point lies between the two.

Before any tool is used, the network is at a tipping point. Six failure pathways are open. Three are Flag and explanation to caseworker, Partner records pooled in HIFIS, and Model retrained on HIFIS records. The other three are Caseworker contacts logged in HIFIS, Caseworker reads client history, and Caseworker weighs the flag's drivers. Two checks do nothing at the start: Standing accuracy check and Second look beyond the shelter data.

This case has a budget of 14 units. Understand the system, which pays for ongoing study of what the deployment is doing, costs 3 units under Explore (No Targets) and Service Targets Only. It costs 4 under the two higher levels, and 6 at its stronger setting at every level. While it is in use, Keep prompts neutral, Upgrade model, Require sign-off, and Assign a challenger each cost 1 unit less, but never less than 1 unit. At its stronger setting they cost 2 units less, with the same floor. Every other tool costs the same at every level.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets are not met before any tool is used. That level asks for the network to be self-correcting and for CHAI to be helping the work.

Three tools meet that level's targets on their own, for 2 units each: Mark AI-written records, Peer sharing rules, and Escalate checks. So do the stronger settings of Mark AI-written records, Peer sharing rules, and Assign a challenger, for 3 units, and Escalate checks at its stronger setting, for 4. Lingering effects is a Dynamics setting, on by default, in which damage outlasts its cause. With it off, Vet connections also does, for 3 units. In all, more than 55,000 combinations of tools and settings within the budget meet them, about 6,500 distinct sets of tools.

Service and Safety Targets and All Governance Targets both ask you to close every failure pathway, among other targets. Closing all six takes at least five tools.

Escalate checks closes Flag and explanation to caseworker. Keep prompts neutral closes Caseworker weighs the flag's drivers. In the Lab it keeps questions put to an AI neutral. CHAI takes no questions, so here it stands for keeping the caseworker's weighing independent of the flag itself.

With lingering effects off, or with Understand the system in use, Vet connections closes Partner records pooled in HIFIS and Model retrained on HIFIS records. Mark AI-written records, or Understand the system, closes Caseworker reads client history. Gate record entries, which requires sign-off before anything enters the records, closes Caseworker contacts logged in HIFIS. Store less data also closes it.

Service and Safety Targets always runs with lingering effects off. There the cheapest sets cost 12 units: Escalate checks, Keep prompts neutral, Vet connections, Mark AI-written records, and either Gate record entries or Store less data. In all, 35 combinations of tools and settings within the budget meet the targets, 20 distinct sets of tools. Every one includes Escalate checks, Keep prompts neutral, and Vet connections. None includes Upgrade model.

All Governance Targets always runs with lingering effects on. Then Vet connections needs Understand the system to close its two pathways. So exactly two sets of tools meet the targets. Both hold Escalate checks, Keep prompts neutral, Vet connections, Store less data, and Understand the system, for 13 units. The second adds Require sign-off, for 14. Keep prompts neutral at its stronger setting is a third way, also for 14.

Gate record entries cannot replace Store less data at that level, because the benefit CHAI adds would then fall below what the level asks. Neither set starts either check.

Using every tool on offer, each at its strongest setting, costs 43 units, about three times the budget. It closes every failure pathway and starts both checks. But the benefit CHAI adds would fall too far, so the targets would be met at no level.