Ottawa has said for years that its software for sorting visa applications doesn’t change how immigration decisions get made. A new study based on more than 17,000 pages of internal government records argues that it does.
Danièle Bélanger and Gabriel Bergevin-Estable published their analysis in the journal AI & Society. They looked at how Immigration, Refugees and Citizenship Canada (IRCC) has brought automated decision support systems, some of them using artificial intelligence, into its handling of temporary resident applications since 2015. Their main case is Chinook, an in-house tool rolled out to visa offices from 2018. The authors conclude that these systems change who exercises discretion over “who gets in,” and also where, when and how they exercise it. In their view, this makes Canada’s administrative borders thicker, more remote and harder to see into.
A quiet legal change, then a wave of tools
The groundwork was laid in 2015. That year Parliament added an “electronic administration” section to the Immigration and Refugee Protection Act, which allows automated systems in decision-making as long as the minister makes them available to officers. When the amendment was debated in the House, MPs raised bias, oversight and the privacy of biometric data. Once it passed, though, IRCC could adopt new systems without any further public or parliamentary debate.
The only other significant rule is the Treasury Board’s Directive on Automated Decision-Making. It took effect in 2019 and requires published Algorithmic Impact Assessments, but only for systems developed or procured after April 1, 2020. The authors note that disclosure was effectively optional from 2015 to 2021. That is why most of their evidence had to come from Access to Information requests filed between January 2021 and March 2024.
IRCC’s stated reason for automating is workload. Its internal figures show that between 2011 and 2019, temporary resident visa applications rose 109 per cent, work permit applications 147 per cent and study permit applications 222 per cent. Over roughly the same period, visa office staff grew by only about 20 per cent.
The records describe a long list of tools. Hiraya was used in the Manila, Colombo and Sydney offices. With it, officers could get through more than 50 files a day, sometimes without opening the attachments, and files were often refused in bulk. An “Advanced Analytics” AI pilot launched in 2018 sorted visitor visa applications from China and India into three tiers. Tier 1 files received automatic eligibility approvals, and officers agreed with the AI on at least 98 per cent of those low-risk cases. The impact assessment for its successor model was not published until January 2022, almost four years after the pilot began. The documents also mention Watchtower, Lighthouse, Starling, Cumulus and a range of office-specific “Officer Rules.” According to the authors, many of these were never disclosed to the public or to Parliament. The impact assessments that were published are self-evaluations, not third-party audits.
What Chinook does
Chinook began as Excel with a Visual Basic layer. It pulls case data from IRCC’s Global Case Management System (GCMS) into a single workbook so officers can act on many applications at once. It has five modules: workload assignment for managers, pre-assessment, decision-making (including bulk decisions), post-decision handling, and a module for risk indicators and “local word flags.” IRCC’s own internal guide describes the tool’s purpose as increasing officer productivity. The department estimates efficiency gains of 18 to 30 per cent.
IRCC has said publicly that Chinook “does not fundamentally change the way applications are processed” and simply offers a “consolidated view” of the file. The study tests that claim and finds several problems.
Risk labels arrive before the officer does
Applications in Chinook may already have been sorted by Advanced Analytics. Tier 1 files receive automatic eligibility approval. Between January 2018 and March 2020, eligibility was the reason given in 65.8 per cent of visitor visa refusals, so an automatic eligibility approval makes a final approval very likely. Files placed in Tiers 2 and 3 get closer scrutiny. According to IRCC documents, those placements can reflect predicted “post-approval adverse events” such as overstays, asylum claims, security threats and crime. The authors compare these components to predictive policing algorithms, a field with well-documented problems of bias.
Applicants are not told which tier their file was placed in or which risk flags were attached to it. Officers are not told why either. The authors argue that this invites automation bias, the tendency of busy people with limited information to defer to a computer’s assessment. They also argue that part of the officers’ discretion moves to the programmers and managers who design the triage rules and flags.
Deciding files in batches
Chinook’s spreadsheet base lets officers sort and filter applications and then decide whole groups at once. The paper gives a hypothetical example. An officer could select every study permit application bound for Quebec and refuse all those missing a Quebec Acceptance Certificate. That would also refuse students on the Canadian Francophonie Scholarship Program, who are exempt from needing the certificate.
Internal documents show that IRCC’s U.S. network of visa offices declined to adopt Chinook’s decision module. Their caseloads were too varied and specialized for grouped refusals to be useful. The authors read this as evidence of where Chinook’s efficiency really comes from: treating applications by what they have in common rather than by their individual details.
Nothing in the immigration law or regulations allows an officer to decide one file based on its similarity to others, the authors note. Chinook also deletes its working data once decisions are exported to GCMS. In Mehrara v. Canada (2024), the Federal Court observed that the “systematic daily deletion of all material generated by processing technology may not reflect best practice.” Without that data, applicants who want to challenge a refusal may have a hard time showing how it was reached.
Documents left unread
Applicants often submit hundreds of pages: bank statements, study plans, job offers, marriage certificates, transcripts. Chinook shows officers structured data from GCMS and the application forms, but not these attached documents. IRCC’s own guidance says the officer is responsible for assessing all of the evidence on file. The records suggest officers may not have enough time to do that, and may rely on summary notes written by other staff or read only enough to reach a decision. Procedural fairness requires that the person who decides a case actually reads what the applicant submitted, so the authors see this as a fairness concern.
Minutes per file and a widening gap
The time spent on each file has dropped sharply. Using Peter Satzewich’s 2010–2012 fieldwork with visa officers, the authors estimate that decisions took between 3 and 12 minutes each. A test of early Chinook processing called Project Peacock recorded 105 decisions per day, which works out to about 3.7 minutes per application. A later time-and-motion review found some Chinook users averaging under 102 seconds per file. Overseas study permit targets assume no more than 7.5 minutes.
Triage has also increased the gap between fast and slow applications. In 2010, the difference between the fastest and slowest applications, up to the 80th percentile of processing times, was at most 20 days. Today it can be more than a year. The factors behind that gap are not disclosed to applicants, to Parliament, or to the officers themselves.
Local knowledge goes missing
Satzewich’s earlier work showed that local expertise was central to good visa decisions, especially the knowledge of locally engaged staff about the country where an application came from. Chinook is designed to standardize procedures across offices and to make it easy to move workloads between them. As a result, a file can be decided by an officer on another continent who has none of that local context. The documents show that officers at secondary offices are told to remove their assignment from GCMS once their work is done. Lawyers have also noticed that refusal letters increasingly leave out which office made the decision. Only the final decision and Chinook’s preformatted refusal notes are carried into GCMS, so there is often no audit trail showing who decided a case or where.
Why the process matters
The authors argue that Canadian administrative law, which mostly looks at the outcome of a decision rather than how it was produced, is poorly suited to this shift. Triage tiers, risk flags and bulk processing shape the discretion officers exercise in ways that matter legally and ethically. They say this calls for a rethink of procedural fairness and accountability when decisions are made with algorithmic help.
The study has limits, which the authors acknowledge. It is based on documents, many of them partly redacted, and contains no interviews with the visa officers who remain legally responsible for each decision. The authors have offered their method of iterative Access to Information requests, catalogued in a searchable database of more than 17,000 pages, as a way for other researchers to study algorithmic systems that governments would rather keep out of view.
Source: Bélanger, D. & Bergevin-Estable, G. (2025). “Revisiting ‘who gets in’: borders and migration management in the era of automation and AI in Canada.” AI & Society, 41, 1089–1102. https://doi.org/10.1007/s00146-025-02602-5 (open access, CC BY 4.0).
