Since 2018, Canada’s immigration department has used machine learning models to sort temporary resident visa applications, now covering every country. In 2023 the Auditor General noted the tool’s effectiveness had not been assessed. The failure isn’t a biased algorithm — it’s eight years of automation nobody has measured.
In January 2022, Immigration, Refugees and Citizenship Canada — the federal department that handles visas, permits and citizenship, known as IRCC — published a technical notice few people read. It said the advanced analytics models used to sort temporary resident visa applications, until then applied only to files from China and India, were being updated, and that a third model would go live covering every other country. From that point on, every visitor visa application submitted to Canada from abroad passes through a machine learning system before it reaches a human officer’s desk.
The artificial intelligence component here is specific and bounded: these are models trained on the historical record of decisions human officers made on comparable applications. They take the structured data of a file as input, produce a classification as output — this one is straightforward, this one needs attention — and in some cases automatically approve the eligibility portion only, meaning the check that an applicant meets the formal requirements of the programme. They do not touch admissibility in the narrow sense: security, criminality, health. And on one point IRCC is categorical: no automated system can refuse an application or recommend a refusal. A refusal is always signed by a person.
What doesn’t exist
Anyone looking for the document that certifies this system’s failure will not find it, because it was never written. There is no Canadian federal audit of automated triage in immigration applications. The Auditor General — Canada’s independent performance watchdog, reporting to Parliament — has produced two reports touching IRCC in recent years: one in October 2023 on permanent residence processing times, and one in March 2026 on the international student programme reforms. Neither has triage as its subject, and neither measures outcome differences by nationality attributable to automation.
The 2023 report is interesting precisely for what it says in passing. Its headline finding concerned waiting: most applicants waited a long time, with refugee programmes the hardest hit. The recommendation asked the department to identify and correct differential wait times. But in the official response, one detail is worth more than many denunciations: the implementation of the automated eligibility assessment tool had not been evaluated for whether it actually reduced processing times.
Read that again. A tool introduced in 2018 to speed up the sorting of files, extended in 2022 to every temporary resident visa application filed from abroad, had not been checked against its stated purpose. This is not an algorithm that discriminates. It is an algorithm whose performance against the reason it was bought has never been measured.
Fifteen million and no meter
The public figure is roughly fifteen million Canadian dollars spent over five years, from fiscal 2017-18 to 2021-22, on advanced analytics activity. That is not a vast sum for a federal department, and that is exactly the point: small enough not to attract a dedicated audit, structural enough to have changed how millions of files are handled.
Meanwhile the perimeter expanded. Spring 2021 brought a model for spouse and common-law partner applications from within Canada, extended in May 2024 to all family class spouse and partner applications. October 2022 added a model for Visitor Records. Late 2023 brought a tool for International Experience Canada work permits, this time accompanied by substantial documentation: an Algorithmic Impact Assessment, peer review, and a GBA Plus assessment — Canada’s required analysis of differential impact by gender and other identity factors. Also in 2023, the Integrity Trends Analysis Tool went live, a machine learning system for spotting risk patterns in applications.
IRCC states that it follows the Treasury Board’s Directive on Automated Decision-Making, Canada’s governing framework for automated administrative decisions, and that it has published more Algorithmic Impact Assessments than any other federal department. That is plausible and probably true. But an Algorithmic Impact Assessment is a document you fill in beforehand: it describes expected risks, planned mitigations, an estimated impact level. It is not a measurement of what the system subsequently did. The gap is the same one that separates an environmental impact statement from emissions monitoring ten years later.
The Chinook case, or an argument about words
Alongside the models sits Chinook, an Excel-based tool in use since 2018, which became public obliquely: through the affidavit of Andie DaPonte, then IRCC’s director of international network optimization and modernization, filed in Federal Court in the litigation over the refusal of Abigail Ocran’s study permit. IRCC maintains that Chinook uses neither artificial intelligence nor advanced analytics, and that it neither makes nor recommends decisions. The Canadian Immigration Lawyers Association contends that with Chinook a single officer can review up to a thousand applications at once — that is the association’s claim, not a departmental figure. In court, the Department of Justice argued the tool “did not substantively affect the outcome”.
The dispute here is not factual but definitional. A spreadsheet that aggregates and reorganises information to speed up human work is not a machine learning model, and calling it AI is technically wrong. But the governance problem it raises is identical: when a technology changes how many files a human eye passes over in a shift, the quality of the decision depends on that technology even if the technology decides nothing. The relevant question is not “who signs” — the answer is known and correct — but “how much human attention is left per file”. None of the available public documents answers it.
The numbers that exist, and what they don’t prove
Canadian refusal rates have risen sharply. According to IRCC parliamentary notes, visitor visa refusals went from 38% in 2023 to 54% in 2024, with 49% recorded in April 2025. Refusals for misrepresentation rose 251% in 2024 over the previous year. Study permit refusals ran at 59% between January and September 2025, against 49% in the same period of 2024.
It is tempting to link that curve to automation. It would be wrong. The department attributes the increase to tighter financial requirements for applicants, the closure of the Student Direct Stream and a sharper focus on programme integrity, and nothing in the available sources establishes a causal link to the triage models. The same applies to the growth in Federal Court applications, reported by secondary sources and disputed by IRCC as regards its attribution to automation.
The courts, for their part, have shut down one argumentative shortcut: according to an analysis published by Law360 Canada in July 2026, the Federal Court has rejected the proposition that the mere use of technology makes a decision unfair or unreasonable, focusing judicial review on the lawfulness of the decision itself. An applicant must show that the specific decision is flawed, not that a computer helped prepare it.
The real failure
The Auditor General’s March 2026 report on the international student programme offers a useful counterexample. The new system for verifying letters of acceptance — not a predictive model but an authenticity check — worked as intended and was used for 97% of more than 841,000 applications. There, measurement exists, and it is positive. The same report finds, however, that the department did not respond effectively to other weaknesses in integrity controls: more than 145,000 students flagged as non-compliant were never verified, and in 800 cases IRCC did not consider acting at all.
This is the portrait of an administration that knows how to build tools and does not know how to close the loop. The failure of Canadian automated triage is not a demonstrated bias or a wrong decision made by a machine: it is the absence, after eight years and three generations of models, of a public answer to the simplest possible question. Did the tool cut processing times? By how much? For whom? The Directive on Automated Decision-Making mandates transparency about expected impact, and IRCC executes that part better than anyone else in the federal government. No rule requires going back to check whether the predictions came true — and that missing piece, not the algorithm, is where the system breaks.