Le triage par apprentissage automatique d'IRCC classe des millions de demandes de visa à l'aide de modèles entraînés sur des décisions historiques

Since April 2018, Immigration, Refugees and Citizenship Canada (IRCC) has used a machine-learning system to triage Temporary Resident Visa (TRV) applications. The system uses IBM SPSS Modeler to generate predictive decision-tree rules from historical immigration decision data, sorting applications into three tiers that determine their processing pathway and materially influence outcomes.

The system has two layers. Layer 1 ("Officer Rules") consists of manually created triage rules developed by IRCC's Beijing visa office using statistical information and historical data. Layer 2 ("Model Rules") is generated by IBM SPSS Modeler, which tests millions of applicant characteristic combinations against historical approval/refusal outcomes to find reliable correlations, then formulates them as decision-tree rules with confidence thresholds.

Applications are sorted into three tiers. Tier 1 applications are classified as "routine" and receive automated eligibility approval with no human review of the eligibility determination — officers only check admissibility (security and criminality). Tier 2 and Tier 3 applications are sent to officers for full review, with Tier 3 carrying the highest refusal rates. The tier designation substantially affects outcomes: Tier 1 applications have near-100% approval rates, while Tier 2 approval rates drop to 63% for online India applications and 37% for India VAC applications. Will Tao, an immigration lawyer who has obtained internal IRCC documents through access-to-information requests, has noted that "Tier 1 Applications are decided with no human in the loop but the computer system will approve them" . IRCC maintains that officers always make the final decision and that the system "never refuses or recommends refusing applications." Tao and other immigration lawyers argue the tier assignment effectively predetermines outcomes even if officers nominally decide.

From April 2018 to January 2022, the system operated exclusively on applications from China and India. This nearly four-year period of nationality-specific ML triage has been identified by researchers and immigration lawyers as the primary discrimination concern. Applicants from these two countries were processed by a machine-learning system trained on historical decisions from those same countries, while applicants from other countries were not subject to algorithmic triage. The model was trained on past officer decisions that may have reflected conscious or unconscious biases. Will Tao's research, based on documents obtained through access-to-information requests, found that historical training guides in Chinese visa offices "assigned character traits and misrepresentation risks based on province of origin." In January 2022, the system was expanded to all overseas TRV applications, and subsequently to Visitor Records and Family Class Spousal applications. IRCC reports that the Advanced Analytics Solutions Centre has assessed more than 7 million applications.

The system was assessed at Level 2 (Moderate) under the Treasury Board's Directive on Automated Decision-Making. Multiple observers have questioned whether this assessment understates the system's impact given its scale and consequences. IRCC published its Algorithmic Impact Assessment on the Open Government Portal in January 2022. A peer review by the National Research Council was conducted in 2018 but was not published until Will Tao obtained it through an ATIP request and published it himself. The Directive's Section 6.3.5 requires peer review publication prior to a system's production; compliance with this requirement has been incomplete.

Applicants are not told which tier they are assigned to. The tier designation is not recorded in GCMS (Global Case Management System) notes. Officers downstream of the triage are reportedly not informed of the rules governing the system. This opacity makes it practically difficult for applicants to challenge a tier assignment they cannot see — though judicial review of the final decision remains available — and officers may not understand what pre-processing shaped the file they are reviewing.

The Canadian Immigration Lawyers Association stated in August 2025 that "the introduction of automated and analytic tools...is directly linked to increase in decisions that are neither meaningful nor well-reasoned." Immigration lawyers have documented patterns of generic refusals, missing document citations for documents that were submitted, and processing timestamps suggesting decisions made in minutes. The AI Monitor for Immigration in Canada and Internationally (AIMICI), founded in October 2025 by Will Tao and three co-founders, was created specifically to investigate and monitor these concerns.

No Federal Court decision has directly addressed the Advanced Analytics triage system. Most litigation has focused on Chinook, a separate data-display tool. In Luk v. Canada (2024 FC 623), the Court held that "the use of algorithms or artificial intelligence to process applications is not in and of itself a breach of procedural fairness." However, in Mehrara v. Canada (2024 FC 1554), Justice Battista noted this "may not be the case in other judicial reviews of applications processed using processing technology, particularly in applications where risk indicators are present" — the first judicial signal that the triage system's impact on high-risk-flagged applications may warrant closer scrutiny.

IRCC describes the system as a triage tool that does not make final decisions — officers retain discretion at every stage, and no application is automatically refused based on tier assignment alone. The department states that the system was designed to improve processing efficiency and reduce wait times. The expansion from two nationalities to global coverage in 2022 addressed the most prominent equity concern about nationality-specific application. The system has been assessed under the federal Directive on Automated Decision-Making, though critics argue the Moderate (Level 2) classification underestimates the system's impact.