Domain Atlas / Housing & homelessness services
CHAI (London, Ontario chronic-homelessness prediction)
London, Ontario's CHAI is a live, caseworker-facing machine-learning model that flags people in the city's shelter system as at risk of chronic homelessness (more than 180 shelter days in a year) about six months ahead; it provides intelligence to prevention caseworkers and does not itself make service decisions. Its widely repeated '93 percent accuracy' is a builder-reported, testing-phase figure from 10-fold cross-validation on historical HIFIS records, never independently validated after deployment; the same technical work reports recall of about 0.921 but precision of only about 0.651, implying substantial false positives under a low base rate.[4]
What happened
The Chronic Homelessness Artificial Intelligence tool (CHAI) is a machine-learning model built by the City of London, Ontario to predict the probability that a person in the city's shelter system will become chronically homeless — defined as more than 180 shelter days within a year — roughly six months ahead. It was built in-house by the City's Municipal Artificial Intelligence Applications Lab within Information Technology Services, with external consultant support for maintenance and periodic (roughly annual) retraining, and the code was published open source under an MIT license. The technical work, released as a September 2020 preprint by Blake VanBerlo and colleagues, describes two architectures — a multilayer-perceptron variant and a hybrid recurrent-plus-feed-forward variant that combines time-series service features with static demographic features — trained on a cohort of 6,521 de-identified individuals structured in 30-day time steps. CHAI draws on de-identified records from the federal Homeless Individuals and Families Information System (HIFIS), pooled across roughly 20 to 24 London homelessness-support organizations; the production system has been described as consuming about 21 million data points, and early use identified 88 people as at risk. During its testing phase CHAI was reported to predict chronic homelessness with "93 percent accuracy"; the underlying paper reports recall of about 0.921 and precision of about 0.651 under 10-fold cross-validation on historical data. Because precision sits well below recall and the base rate is low, "accuracy" overstates operational reliability — the model produces substantial false positives for each true chronic case — and no independent post-deployment validation of predictive accuracy or downstream outcomes was found. CHAI went live in August 2020 during the city's COVID-19 response; the build cost roughly CA$57,000 with ongoing costs near CA$1,100 per month, and city materials framed it as providing intelligence to caseworkers rather than making decisions itself. Two design choices are distinctive and genuine. It is consent-driven: it uses information from people who are in HIFIS and lets individuals opt out of having their data included, and it was built with reference to GDPR principles and Canada's Directive on Automated Decision-Making. And it was designed for interpretability, using local feature-attribution explanations so caseworkers can see which features drove a prediction, an approach the builders framed as building trust and reducing unintended bias rather than operating as a black box. The city justified the tool with cost figures holding that a chronically homeless individual costs roughly CA$135,000 per year and uses about 534 shelter days versus about 45 for a non-chronic user, roughly a twelve-fold difference. Independent scrutiny has followed. A review commissioned by Global News in 2023 — two unaffiliated computer-science experts and a privacy lawyer — found CHAI appeared to take the steps needed to protect personal information, but reviewers noted it remains "a model of surveillance" and that HIFIS-based data can miss vulnerable groups, listing aging people, women with children, new immigrants, Indigenous families and users of private shelter systems. A CBC News report at launch and the Global News reviewers both raised the point that a public-shelter definition of homelessness leaves out people who couch-surf or make private arrangements, disproportionately women. An ongoing Western University / Starling Centre research project situates CHAI in its historical, legal and policy context and raises fairness, accountability, surveillance and inequality concerns. By 2024, as Ottawa and other cities adopted similar tools, London was described as the Canadian pioneer of AI chronic-homelessness prediction, and coverage surfaced ethical concerns that raising a person's risk score for identifying as Indigenous is problematic and that predictive systems can obscure the structural root causes of homelessness and take pressure off policy-makers; the president of the Canadian Alliance to End Homelessness stressed that the key to ending homelessness is housing.
The sociotechnical reading
CHAI is the rare case that does much of the visible governance well and still turns on something the governance never touches. It explains itself to the caseworker instead of hiding behind a score, it lets people opt out, it was published open-source and reviewed for privacy, and it never makes the service decision itself. Almost every critique the other predictive tools in this Atlas attract — opacity, no human in the loop, no consent — CHAI has an answer for. And yet the decisive question sits upstream of all of it: who can be scored at all. A model that runs on public-shelter records can only ever see people who come through a public shelter door. The many who couch-surf, double up, make private arrangements, or avoid the shelter system — disproportionately women, families, new immigrants, and some Indigenous people — are not scored badly; they are absent, invisible to a tool that cannot represent a person its data never contains. The consent design, a real legitimacy gain, tightens this frame rather than widening it: an opt-out is a genuine right, but every opt-out removes a person from the population the model can see, and the opt-out itself is invisible to the model. That is the distinctive lesson of this case. The governance object here is not the classifier's accuracy or even the human's discretion; it is the sampling frame — the front door that decides who is counted before the model runs — and consent-based inclusion is double-edged, trading representational coverage for legitimacy in a way no amount of downstream accuracy, interpretability, or human review can undo. The contested headline compounds the point. The "93 percent accuracy" that travels with CHAI was a testing-phase cross-validation figure on historical data, and the same technical work reports precision far below recall: many false positives for each true chronic case under a low base rate. It was never independently re-validated after deployment, and the one independent review that happened examined data protection, not whether the flags still hold up. In map terms, the levers that matter are the ones that govern the frame and the evidence, not another model upgrade: keep a standing check on whether the flags still predict what they claim; make the headline number re-earn its place under independent re-validation before anyone repeats it; label the retraining loop so the model's own past flags do not silently become tomorrow's ground truth; and, hardest of all, build a routine second look at who never entered the data at all. The honest reading keeps the benefit in view — this is proactive prevention with real consent and real explanation, not a punitive scoring machine — while naming the trap: a tool can be transparent, consensual, and accurate on its own terms and still quietly encode who a city is able to see.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.