Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

CHAI (chronic-homelessness prediction)

The people the data can't see: a consent-based homelessness-risk model

A city built its own model to flag shelter clients heading toward chronic homelessness, and did much of it right: it explains its reasoning to the caseworker, it lets people opt out, and it never makes the call itself. Modeled on London, Ontario's CHAI. The catch is upstream of all of that: the model can only score people who show up in the public-shelter data, and its widely repeated accuracy was a testing-phase number no one ever checked again. Watch who never appears on this diagram at all.

Stylized model of a documented deploymentHousing & homelessness services

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the CHAI-class chronic-homelessness risk model network: 6 components and 12 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 4 assumed · 2 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    This models the consent-based, caseworker-facing predictive-flagging pattern documented in the CHAI (City of London, Ontario) case file — not a reconstruction of the actual model or its code.

  • baseline

    The defining upstream dynamic is selection at the front door: only people who appear in public-shelter HIFIS records can be scored, and the opt-out self-selects the population further, so the modeled risk surface is a selected sample of actual need. Who is structurally missing — many women, families, new immigrants, some Indigenous people, and private-shelter users — is where the harm lands, and it is recorded in the case file, not computed here.

  • assumed

    The consent-gated intake and the retraining loop are both drawn privacy-sensitive: ~20-24 partner agencies pool de-identified intake records into the one shared HIFIS record system (a store-to-store pooling pathway, the shape a connection-authorization lever governs), and the accumulating record is periodically retrained on, so the model's own flags feed back into what it later learns. What that reuse means for the people in the shelter data is recorded externally in the case file, never computed in these dynamics.

  • baseline

    Unlike the opaque-score cases, the caseworker-to-model channel is drawn as real and load-bearing: CHAI was designed for interpretability, so the caseworker can see which features drove a flag and weigh it. That interpretability is a genuine governance strength, and it is why the human-in-the-loop discretion here is more than nominal.

  • assumed

    Both peer-check pathways start closed: there is no standing independent post-deployment audit of the flags' accuracy or outcomes (the commissioned review examined data protection only), and there is no routine second look at who never entered the shelter data at all. These are the latent check pathways levers can open; the case file explains why no independent accuracy validation was found.

  • assumed

    The widely repeated 93 percent accuracy is a builder-reported, testing-phase figure from a 10-fold cross-validation on historical records, never independently validated after deployment; the same technical work reports precision far below recall, implying many false positives under a low base rate. This Lab models institutional propagation only and estimates none of those figures. Chronic homelessness itself, and differential harm to the people the model can or cannot see, are not modeled here.

What this example does not show

  • The people this system flags — shelter clients — are not in this diagram, and neither are the people it cannot see. The Lab models how a flag moves through the city's own caseworker workflow, never who becomes chronically homeless or is prevented from it. Who is structurally missing from the public-shelter data, and any pattern in who gets flagged, are recorded in the case file and measured outside a diagram like this one.
  • The widely repeated 93 percent accuracy is a builder-reported figure from a testing-phase cross-validation on historical records, never independently validated after deployment; the same technical work reports precision far below recall, implying many false positives under a low base rate. Nothing in this Lab estimates or endorses that figure, and no clinical or housing outcome for any person is modeled here.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • CHAI is consent-based: it draws on de-identified HIFIS records pooled from roughly 20 to 24 London homelessness-support organizations and lets individuals opt out of inclusion, and it was built with reference to GDPR principles, Canada's Directive on Automated Decision-Making, and local feature-attribution explanations for caseworkers. Because HIFIS captures people who use public shelters, an independent review and reporting at launch note it can under-represent or miss groups who avoid them - including many women, families, new immigrants, some Indigenous people, and private-shelter users; academic researchers situating the tool raise related fairness and inequality concerns. So the population the model can score is a selected sample of actual need, and the opt-out self-selects it further.

    empirical
    • Trade press Wray, Explainable AI Predicts Homelessness in Ontario City (Cities Today, 2020) https://cities-today.com/explainable-ai-predicts-homelessness-in-ontario-city/
    • Investigative LeBel, How One Ontario City Is Blazing the Trail for Public Sector AI Use (Global News, 2023) https://globalnews.ca/news/9765050/london-ontario-artificial-intelligence-homelessness/
    • Investigative Lamberink, A City Plagued by Homelessness Builds AI Tool to Predict Who's at Risk (CBC News London, 2020) https://www.cbc.ca/news/canada/london/artificial-intelligence-london-1.5684788
    • Academic Redden, Stark, Centivany, Lizotte, Adler, Situating London's AI Homelessness Model (Starling Centre for Just Technologies, Just Societies, Western University, ongoing; accessed 2026) https://starlingcentre.ca/project/situating-londons-ai-homelessness-model/

  • 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.

    empirical
    • Trade press Wray, Explainable AI Predicts Homelessness in Ontario City (Cities Today, 2020) https://cities-today.com/explainable-ai-predicts-homelessness-in-ontario-city/
    • Academic VanBerlo, Ross, Rivard, Booker, Interpretable Machine Learning Approaches to Prediction of Chronic Homelessness (arXiv:2009.09072 preprint, 2020) https://arxiv.org/abs/2009.09072
    • Investigative LeBel, How One Ontario City Is Blazing the Trail for Public Sector AI Use (Global News, 2023) https://globalnews.ca/news/9765050/london-ontario-artificial-intelligence-homelessness/
    • Trade press Govlaunch Stories, London, ON Uses AI to Fight Chronic Homelessness (2020) https://govlaunch.com/stories/london-on-uses-ai-to-fight-chronic-homelessness

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Housing & homelessness services domain page.

Levers available here and the patterns behind them

Documented case histories