By Stuart Kerr, Technology Correspondent, LiveAIWire
AI at the border is not a future scenario. It is already deciding, in whole or in part, whether specific named people get to enter a country, keep a visa, or stay out of detention, and the record of what happens when those systems get it wrong is now long enough to draw conclusions from.
In 2020, the UK Home Office agreed to scrap its visa “streaming” algorithm after the campaign group Foxglove and the Joint Council for the Welfare of Immigrants threatened a judicial review, in what became the first successful judicial review of a UK government algorithmic decision-making system. The tool had assigned a secret “red, amber, green” risk score partly on the basis of nationality, and had been used to sort every UK visa application for years.
That case established the template for almost every AI at the border controversy since: a government adopts an automated tool to manage volume, the tool encodes an existing bias or introduces a new error, and the people harmed are the ones with the least institutional power to contest it before the harm is done.
The Case That Shows Exactly How AI at the Border Goes Wrong
Canada’s Immigration, Refugees and Citizenship Canada has used a tool called Chinook since 2018 to help process high volumes of visa and residency applications, generating standardised language that officers can select, modify or reject when writing refusal decisions. In 2026, a French immunologist with a PhD from Sorbonne University, identified in reporting as Kémy Adé, received a permanent residency refusal that described a job she had never held, with duties that bore no relationship to her actual research career. The refusal letter’s fabricated content became, in the words of immigration lawyers tracking the case, exactly the kind of documented, concrete discrepancy that Canada’s Federal Court has said is required to challenge an AI-assisted decision.
That evidentiary bar exists because of how the Federal Court has actually ruled on AI at the border in Canada so far. In Haghshenas v Canada (2023), the court held that the use of Chinook is not itself a procedural fairness problem, provided a human officer renders the actual decision. In Espinosa Cotacachi v Canada (2024) and Pjetracaj v Canada (2025), the court went further, ruling that speculation about how the tool works is not enough.
An applicant needs clear, documented evidence, like a refusal letter that invents a job that never existed, to show that a human reviewer did not meaningfully engage with their file. The effect of that standard is that the burden of proving an algorithm caused a wrongful decision falls almost entirely on the person the decision was made about, using evidence they usually only see after the damage is done.
The US Has Moved From Sorting Applications to Tracking People
The United States has taken AI at the border further than either the UK or Canadian examples, extending it from processing applications to tracking people after they are already inside the country. Immigration and Customs Enforcement is deploying a Palantir-built system called ImmigrationOS, which the American Immigration Council has described as designed to track immigrants’ movements and consolidate data across agencies for enforcement purposes.
The Department of Homeland Security has also incorporated AI into US Citizenship and Immigration Services’ assessment of benefit eligibility and asylum credibility, and into ICE’s electronic monitoring decisions, according to a September 2024 letter from the Electronic Frontier Foundation and more than 140 other civil society organisations calling on DHS to stop using AI in immigration decisions entirely. Their central argument was not that the technology is imperfect in the ordinary sense that all software is imperfect. It was that immigration law is complex enough, and the consequences of error severe enough, including detention and wrongful deportation, that even a low error rate produces outsized harm concentrated on people with limited legal recourse.
Why the Law Has Not Caught Up
The European Union has gone furthest of any jurisdiction in writing specific rules for AI at the border. Under the EU AI Act, AI used in migration, asylum and border control management is classified as high-risk under Annex III, triggering mandatory risk management, human oversight and fundamental rights impact assessments. Recital 60 explicitly states that such systems must never be used to circumvent the 1951 Refugee Convention or the principle of non-refoulement, the prohibition on returning someone to a country where they face serious harm.
In practice, legal scholars examining the Act’s border-specific provisions have identified a significant gap: predictive systems that forecast migration movements were nearly excluded from the high-risk category entirely, and the compliance mechanism for border authorities relies on internal self-assessment rather than independent, external audit. The political agreement to push the AI Act’s high-risk compliance deadline for this category from August 2026 to December 2027 has not yet been formally adopted, which means the current, earlier deadline remains the legally operative one for now, but the direction of travel is toward more time for authorities to comply, not less.
Our own reporting on whether algorithms can predict refugee displacement found the same tension at the humanitarian end of this technology: forecasting tools built to help aid agencies prepare for displacement rely on exactly the kind of passive data collection, satellite imagery, border scans, aid distribution patterns, that refugees never consented to and cannot meaningfully contest. Our coverage of how AI is being used to track trafficking and smuggling networks found the same facial recognition and pattern-matching tools built to protect trafficking victims can misidentify people or enable surveillance of vulnerable populations when deployed without transparency, a risk that applies with equal force when the same categories of tool are pointed at migrants and asylum seekers rather than traffickers.
What Every AI at the Border Case Has in Common
What connects the UK’s scrapped algorithm, Canada’s Chinook litigation, and the US enforcement build-out is not that any single tool is uniquely dangerous. It is that AI at the border operates in one of the few areas of public administration where the person affected by an automated decision has the least ability to see how it was made, the least time to challenge it, and the most to lose if the challenge fails.
A wrongly denied loan can be appealed with the borrower still living in their own home. A wrongly refused visa can mean a cancelled research position, a family kept apart, or a return to danger, often before there is any meaningful opportunity for a human to double-check what AI at the border actually decided.
The evidence to date does not support removing AI at the border from immigration processing altogether, since these systems exist because human caseworkers cannot process current volumes without them. It does support the specific, narrow standard that Canada’s Federal Court, the EU’s AI Act, and the UK’s own scrapped algorithm all converge on from different directions: a human being must meaningfully review the specific facts of each case, and the applicant must be able to see enough of the reasoning to identify a documented error.
The system cannot be allowed to become the actual decision-maker in practice merely because a human’s name is on the final letter. Our own reporting on algorithmic risk scoring in sentencing and parole decisions found the same structural failure in a different corner of government: predictive tools that shape decisions about a person’s liberty, reviewed by officials who are told the algorithm’s output but rarely shown enough of its reasoning to meaningfully second-guess it. AI at the border and AI in the courtroom turn out to be the same accountability problem wearing two different uniforms.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.