Immigration paperwork is already a slow grind of forms, deadlines, and decisions that can rearrange a person's entire life. Kémy Adé did the grind. She is a health scientist with a PhD from Sorbonne University in the immunology of ageing, and she had been in Canada since 2023, working as a post-doctoral research fellow and guest teacher at McMaster University. Then Immigration, Refugees and Citizenship Canada (IRCC) refused her application for permanent residence, and the reason it gave described a completely different person.

According to the Toronto Star, which broke the story, with follow-up coverage in the Waterloo Region Record and the Jamaica Gleaner, the refusal letter said Adé's job duties included "wiring and assembling control circuits, building control and robot panels, programming and troubleshooting," and that these did not match the Canadian work experience she had claimed. IRCC was right that the duties did not match. It was wrong about whose duties they were. Adé does not wire control circuits or build robot panels. She studies the immune system. None of those tasks appeared anywhere in the application she had filed roughly a year earlier.

A refusal built on duties she never had

Read the letter cold and the logic almost makes sense: the applicant's stated work experience does not line up with the duties on file, so the claim fails. That is a normal, defensible reason to refuse an application. It only collapses when you notice that the "duties on file" were not hers. The mismatch the system detected was a mismatch it had manufactured. It compared her real qualifications against a fictional job and, unsurprisingly, found a discrepancy.

Adé described being disoriented by the letter. That is a measured response to reading an official document that confidently attributes an industrial-controls career to a biomedical researcher. The harm here is not abstract. A permanent residence refusal can affect a person's right to stay, work, and build a life in the country where they already live with their family. Adé came to Canada in 2023 on a work permit. A refusal grounded in invented facts is not a minor clerical slip; it is a decision with real weight, justified by something that never happened.

A disclaimer that said the quiet part out loud

What makes this case more than a one-off error is the fine print. At the bottom of the refusal letter sat a disclaimer noting that generative AI had been used to support processing of the application. The same disclaimer insisted that all generated content was verified by an officer and that generative AI was not used to make or recommend the decision.

Immigration lawyers tracking the case believe it is the first time IRCC has explicitly acknowledged generative AI in the context of an immigration refusal. The timing was almost too on the nose: Adé's refusal landed in late February 2026, right as IRCC published its first formal AI strategy, a document promising to use artificial intelligence to "boost efficiency, enhance service delivery and strengthen program integrity." The strategy states plainly that AI tools do not refuse or recommend refusing applications, and that human officers make every refusal based on their own review.

So the official position is reassuring. AI assists; humans decide. The problem is that the assistance, in this instance, appears to have fabricated the central fact the decision rested on.

Human in the loop, or human as rubber stamp

Adé's lawyer, Luka Vukelic, went straight at the weakest point in IRCC's defense. "I cannot comprehend how any human being could make this decision," he said. "Somehow, it hallucinated my client's job description. I would love to see what the officer saw. Something seriously went wrong here."

That is the right question. "Verified by a human officer" is supposed to be the safeguard that catches exactly this kind of nonsense. A person reads the generated content, checks it against the actual file, and corrects anything that does not belong. If that review had functioned as advertised, the phrase "building control and robot panels" should have stopped any officer cold, because it bears no relationship to an application about immunology research.

Either the officer did not meaningfully compare the generated summary against the underlying application, or the review process is structured so that the AI output carries more authority than the file it is supposed to summarize. Both possibilities are bad. The first means "human in the loop" was theater. The second means the loop is wired backwards, with the machine's confident hallucination treated as the baseline and the human reduced to a formality. A safeguard that approves a fabricated job description is not a safeguard. It is a signature.

This is the recurring failure mode whenever AI gets dropped into high-volume bureaucratic work. The tool is sold as a time-saver for overloaded staff, and the entire value proposition depends on humans not spending much time on each item. Then the one task that actually required careful human attention, catching the fabrication, is the task the time-saving pressure quietly eliminates.

Why a hallucinated job description matters here

Generative AI hallucination is a well-documented behavior. These models produce fluent, plausible text whether or not it is grounded in anything real, and they do not signal the difference. In a chatbot answering trivia, a confident fabrication is annoying. In a document that determines someone's legal status, the same behavior becomes a fairness problem with a named victim.

IRCC has been leaning on automation for years. It has run an internal data-analytics tool known as Chinook and has explored machine-learning systems to triage and sort an immigration caseload that now exceeds a million applications. The pitch is always efficiency, and the backlog is real. But efficiency that introduces fabricated facts into refusal letters is not efficiency; it is a faster way to be wrong, with the added cost that the errors are dressed in official language and harder to challenge.

There is also a transparency problem. Applicants and their lawyers describe the system as a black box. If a refusal is partly shaped by generated content, the affected person needs to know what the model produced, what the officer actually reviewed, and how the conclusion was reached. Without that, an applicant cannot meaningfully contest a decision; they are left arguing against an output nobody will fully show them.

What reopening the file does and does not fix

Adé's lawyer requested a reconsideration, and IRCC reopened the file. According to the Canadian AI Case Tracker maintained by Anand & Siu LLP, the matter was filed at the Federal Court as IMM-5090-26, and on May 1, 2026 the application was discontinued after IRCC accepted the reconsideration request and reopened the case. So the immediate damage to one applicant is on a path to being undone.

That is genuinely good news for Adé, and it is the system working the way an appeals process is supposed to. But a reconsideration is a patch, not a fix. It depends on the applicant noticing the error, having the resources to retain a lawyer, and acting inside short judicial-review deadlines. Plenty of applicants have none of those advantages. For every refusal challenged by a determined lawyer, there are others who simply accept a decision they assume the government got right, never knowing that the "facts" against them were assembled by a model that does not know what is true.

The lesson is not that IRCC should never touch AI. Automation can genuinely help with sorting, summarizing, and surfacing information across a brutal backlog. The lesson is that when a system generates content that ends up in a legally binding decision, the verification step has to be real, documented, and accountable, not a box-ticking ritual. If an officer signs off on a hallucinated job description, the agency owns that signature. A disclaimer at the bottom of the page does not transfer the responsibility to the software.

Canada asked AI to make immigration processing faster. In this case it got a refusal letter about a scientist who builds robot panels for a living, which she does not, and a reminder that "verified by a human" is only worth something if a human actually looked.