The Immigration Algorithm: Lives and Borders

In October 2022, Canada’s Immigration, Refugees and Citizenship Canada acknowledged that it had been using an AI-based “Chinook” system to assist in processing immigration and refugee applications since 2018. The system was not disclosed to applicants. Its role in decision-making was not explained in refusal letters. Applicants who received negative decisions had no means of knowing whether an algorithm had been involved or what weight it had been given. This was disclosed not through transparency publication but through Access to Information requests by researchers at the Global Migration Lab.

The pattern is global and consistent. Immigration is the domain where government use of AI is most consequential (the decisions determine where people can live, whether families are separated, whether people fleeing violence receive protection) and most opaque (the systems are least disclosed, least subject to independent audit, and least constrained by administrative transparency requirements that apply in other domains).

This is not accidental.

The immigration exceptionalism problem

Immigration law in most countries contains exceptional provisions that reduce procedural protections relative to other administrative decisions. The legal doctrine of plenary power in the United States, which grants Congress and the executive branch broad discretion over immigration decisions with limited judicial review, creates a zone where due process requirements are attenuated. In the UK, the Immigration Rules are secondary legislation that can be changed rapidly without primary legislative process and that contain provisions allowing for expedited decisions without the procedural steps required in, say, social security determinations. In Australia, the Minister for Home Affairs has personal discretion powers that operate outside the standard merits review system.

This means that when AI systems are deployed in immigration, they operate in an institutional environment with fewer of the accountability mechanisms that constrain AI use in other government domains. The right to a human reviewer, the right to an explanation, the right to challenge a decision on procedural grounds — these are all weaker in immigration than in benefits, or tax, or licensing.

What this creates is a high-consequence, low-accountability environment — precisely the worst combination for AI deployment. The decisions matter enormously. The redress mechanisms are limited. The public visibility is restricted by confidentiality provisions that apply to immigration proceedings. The populations affected (migrants, asylum seekers, visa applicants) have limited political voice in the countries making decisions about them.

The visa streaming system

The UK Home Office operated an algorithmic “streaming” system for visa applications from 2015 to 2020. The system sorted applications into traffic-light risk categories — red, amber, green — which determined the level of scrutiny applied. Green applications received minimal examination; red applications received intensive scrutiny.

The system used nationality as a primary sorting variable. Applications from nationals of certain countries, regardless of individual circumstances, were systematically routed to higher scrutiny levels. A leaked Home Office document reviewed by the Bureau of Investigative Journalism showed that the nationality rankings embedded in the system tracked racial and socioeconomic classifications that would have been directly discriminatory if applied by a human caseworker.

When the Joint Council for the Welfare of Immigrants challenged the system in 2020, the Home Office discontinued it rather than defend it in court. The decision to discontinue rather than disclose what the system was actually doing has meant that the full extent of discriminatory impact has never been publicly quantified. Thousands of applications were assessed under a system that the Home Office itself could not defend, and the people whose applications it affected have no way of knowing whether nationality-based routing affected their case.

US CBP and the algorithmic border

US Customs and Border Protection has deployed a suite of AI systems at border crossings and airports, across immigration pre-screening, document verification, facial recognition, and behavioral analysis. The Automated Biometric Identification System (IDENT), the Traveler Verification Service, and various risk-scoring systems work at different points in the journey — some before travel, some at the point of entry.

The oversight of these systems is fragmentary. CBP operates with a congressional mandate that prioritizes security over procedural niceties and a budget that has grown substantially (the agency’s technology procurement budget increased from $1.7 billion in 2019 to $3.2 billion in 2026). The Privacy Impact Assessments that CBP publishes for its biometric and AI systems describe what the systems do and what data they collect, but do not include independent accuracy validation, demographic performance breakdowns, or systematic analysis of false positive rates by nationality, race, or other characteristics.

The practical consequences of false positives in border screening are severe. A traveler flagged as a security risk may be denied boarding, detained at entry, or placed in secondary screening that can last hours and result in visa cancellation. There is no administrative appeal mechanism for real-time border screening decisions. A CBP officer’s decision, even when driven substantially by algorithmic flagging, has extremely limited review. If the algorithm is wrong about you at an airport, your options are to miss your flight while the matter is resolved, or to return home.

Academics at Georgetown Law’s Center on Privacy and Technology have filed more than 30 Freedom of Information requests over four years seeking CBP’s internal validation data for its biometric systems. Most requests have been denied in full or partially redacted under law enforcement exemptions. The accuracy claims CBP makes in public (generally very high accuracy rates for document verification, lower but still high for facial recognition) come from the agency’s own reporting with no independent verification methodology.

Refugee determination and the pretense of objectivity

The most consequential immigration AI applications are in refugee status determination — the assessment of whether someone has a well-founded fear of persecution that entitles them to international protection. These decisions are explicitly required by law (the 1951 Refugee Convention and its 1967 Protocol) to be individual, based on the specific circumstances of each case, and subject to fair procedure. They are also expensive, time-consuming, and backlogged everywhere that receives substantial asylum claims.

Several countries, including the Netherlands, Germany, and Australia, have piloted AI tools to assist asylum caseworkers. The tools vary: some are document verification systems (flagging potentially fraudulent identity documents), some are country condition information retrieval systems, some are case management aids that identify similar precedent decisions. The policy discourse around these tools is careful to describe them as “decision support” rather than decision-making.

The distinction between support and decision-making erodes under caseload pressure. When a caseworker handling 800 cases per year has an AI tool that suggests a case is low-risk and similar to a set of previous denials, the path of least resistance is to follow the suggestion. The tool influences the outcome without being formally responsible for it — a structure that is convenient for both the vendor (not liable for the decision) and the agency (the decision remains technically human) but provides no accountability for the algorithm’s systematic errors.

Human Rights Watch documented in 2024 that asylum seekers whose cases were processed using AI-assisted tools in Germany received different outcomes from those processed without AI tools for similar factual profiles — not because the tool was determining the outcome, but because caseworkers in tool-assisted workflows spent significantly less time per case and were significantly less likely to probe inconsistencies in the formal record. The AI was improving throughput and reducing the quality of individual assessment simultaneously, in ways the statistics couldn’t distinguish.

The accountability gap in immigration AI is solvable. It requires transparency about which systems are used, demographic breakdown of outcomes for algorithmically-assisted decisions, genuine independence for oversight bodies, and legal rights of appeal that include the ability to challenge algorithmic inputs. These are not technically demanding requirements. They require the political will to apply to immigration the same standards applied elsewhere. Given that the populations affected are not voters in the countries whose governments are making these choices, the political will is harder to mobilize than the problem is to solve.