IMMERGITY
IRCC Used Generative AI — And It Invented an Applicant's Job Description
A post-doctoral researcher at McMaster University had her PR application refused after IRCC's generative AI system fabricated her job duties as an electrician. The refusal letter contained the first known explicit AI disclaimer in Canadian immigration history — and the case exposes a fundamental accountability gap in how IRCC deploys these tools.
Part of the IMMERGITY IRCC Errors & Reconsideration Canada 2026 →
At a recent gathering of immigration practitioners, a colleague raised a case that stopped the conversation cold. A post-doctoral research fellow at McMaster University — a health scientist with a PhD from the Sorbonne, specialising in the immunology of aging — had her permanent residence application refused. The refusal letter described her job duties as wiring and assembling control circuits, building control and robot panels, and programming and troubleshooting industrial equipment.
She is not an electrician. She has never been an electrician. None of those duties appear anywhere in her application file. What the refusal letter did contain — buried at the bottom — was a disclaimer that the decision was supported by generative AI.
The case of Kémy Adé, reported by the Toronto Star on March 25, 2026, is the first documented instance of IRCC explicitly disclosing the use of generative AI in an immigration refusal. Her lawyer, Luka Vukelic, submitted a reconsideration request, and the file was subsequently reopened. But the damage — months of status uncertainty, the cost of legal intervention, and the psychological toll on a person who followed every rule correctly — is not undone by a reopened file.
What the Refusal Letter Actually Said
Adé's PR refusal arrived in late February 2026, citing her current job duties as inconsistent with the Canadian work experience she had claimed. The duties listed bore no resemblance to the role she described in her application. Her actual work — health science research, teaching, and post-doctoral fellowship activity at McMaster — was nowhere in the officer's reasoning.
The disclaimer at the bottom of the letter read, in substance: generative AI was used to support application processing; all generated content was verified by a human officer; generative AI was not used to make or recommend a decision.
That disclaimer created an immediate contradiction. If a human officer verified the AI-generated content, and the content described a health scientist as an electrician, one of two things happened: the officer did not read the original documents, or the officer read them and still approved the AI output. Neither scenario reflects acceptable administrative practice.
Why Generative AI Behaves Differently From Earlier Automation
IRCC has used automated tools since 2013. Earlier systems — including the Chinook platform and the Integrity Trends Analysis Tool — work by classifying, sorting, and flagging applications based on predefined rules and pattern matching against existing datasets. They do not create new content.
Generative AI operates on a fundamentally different basis. It uses large language models to synthesise new text by applying learned patterns to a given prompt. As Toronto immigration lawyer Zeynab Ziaie, co-founder of AI Monitor for Immigration in Canada and Internationally, explained in the Toronto Star reporting: You give it a prompt and it can use its large language models to create that response for you and build on what your prompt is to give you a refusal letter. Or it could give you on the same prompt an acceptance.
The critical word is hallucination — when a language model generates plausible-sounding but factually incorrect content, drawing from unrelated training data. An occupational description for an industrial electrician (NOC 72200) shares structural characteristics with other skilled trade and technical roles. When prompted to summarise a file's work experience, a model trained on millions of job descriptions may conflate an unrelated NOC with the applicant's actual role, particularly if the file summary prompt is ambiguous or the model retrieves from a broad context window.
This is not speculation. It is what the Adé case documents: an AI system generated a job description that matched no document in her file, and that output was placed in an official refusal letter bearing IRCC's authority.
IRCC's AI Strategy — Published the Same Month as This Refusal
The timing of this case is not incidental. IRCC published its first-ever Artificial Intelligence Strategy in February 2026 — the same month Adé's application was refused. The strategy states that IRCC will avoid autonomous AI systems that can refuse applications, that human verification will remain central to all decisions, and that AI will be used to support efficiency in areas such as triage, summarisation, fraud detection, and email management.
The strategy also acknowledges that IRCC is currently experimenting with publicly available generative AI tools. It does not specify which tools, what prompts are used, how outputs are reviewed, or what quality controls exist to catch hallucinated content before it enters a decision.
McGill University law professor Jennifer Raso, an expert on digital government and administrative decision-making, raised the questions the department has not answered: At what point was Gen AI used? Are we talking about summarising documents? Does the officer look at those original documents at all? Are those documents translated by Gen AI or not? IRCC did not respond to the Toronto Star's questions on what tools are used, how they are tasked, or what verification involves.
The Legal Accountability Problem
Under Canadian administrative law, a person has the right to know the reasons for a decision that affects them. That principle underlies the entire IRCC reconsideration and judicial review framework. The reasons must be intelligible, transparent, and justified — the standard set by the Supreme Court of Canada in Dunsmuir and reinforced in Vavilov (2019).
When a refusal letter cites fabricated job duties as the basis for finding that an applicant's work experience does not meet requirements, the reasons fail on their face. The duties cited do not appear in the file. They cannot have been the product of a reasonable assessment of the evidence. They are, by any fair reading, not the applicant's duties at all.
The disclaimer that a human officer verified the AI output does not resolve this. Verification that produces a result contradicted by every document in the file is not meaningful verification. And IRCC's simultaneous claim — that the human officer made the decision and generative AI played no role — is directly contradicted by the disclaimer on the same letter.
Practitioners in this space will recognise what this means practically: an applicant facing a refusal based on AI-hallucinated content cannot know, without a GCMS notes request, whether the officer actually reviewed their original documents or relied on a generated summary. That opacity undermines the right to procedural fairness.
What Happened to Adé's Application
After the case received public attention through the Toronto Star, Adé's lawyer submitted a formal reconsideration request. IRCC reopened the file. As of the updated article date (May 6, 2026), the application was back under active assessment.
That outcome is the best available resolution within the reconsideration framework. But it required: a Toronto Star investigation, public pressure, experienced legal representation, and the willingness of a named applicant to speak publicly about a decision that caused significant disruption to her life and career at McMaster.
Most applicants in this situation have none of those resources. They receive a refusal, see cited duties that bear no relation to their work, and have no framework for understanding whether the error originated in AI output, officer negligence, or a file mix-up. The reconsideration process is not a guaranteed remedy — as discussed in IRCC Refused Your Application by Mistake?, reconsideration is informal, unregulated, and entirely at the officer's discretion.
How IRCC's AI Tools Compare: Earlier Automation vs. Generative AI
| Tool Type | How It Works | Can It Hallucinate? | Role in Decisions |
|---|---|---|---|
| Chinook (workflow automation) | Routes and organises applications by predefined rules | No — rule-based only | Triage and assignment |
| Integrity Trends Analysis Tool | Pattern-matches against datasets of 1.4M+ applications to flag fraud indicators | No — classification only | Integrity flagging |
| Generative AI (LLM-based) | Synthesises new text from patterns learned across training data | Yes — documented property of all LLMs | Summarisation, research support (as of Feb 2026) |
Key Timeline: The Adé Case and IRCC's AI Strategy
| Date | Event | Significance |
|---|---|---|
| February 2026 | IRCC publishes first-ever AI Strategy | First formal disclosure of AI use in immigration processing |
| Late February 2026 | Kémy Adé's PR application refused | Refusal cites electrician duties; AI disclaimer appears on letter |
| March 25, 2026 | Toronto Star breaks the story | First public reporting on IRCC generative AI hallucination in a refusal |
| May 6, 2026 | Toronto Star updates article | File confirmed reopened following lawyer's reconsideration request |
Your Response Options When a Refusal Cites Wrong Information
| Option | Deadline | Reliability | When to Use |
|---|---|---|---|
| Reconsideration request (webform) | No fixed deadline — but act immediately | Informal; officer has full discretion to ignore | Clear factual error in file; officer error confirmed by GCMS notes |
| Judicial Review (Federal Court) | 15 days from refusal (inside Canada); 60 days (outside Canada) | Most reliable for clear procedural errors | Refusal reasons are unreasonable on their face; reconsideration ignored or denied |
| Reapplication | No deadline | High — starts fresh with corrected file | Parallel to JR/reconsideration; viable if status permits |
| MP inquiry | No deadline | Can accelerate reconsideration; not a decision mechanism | Supplement to reconsideration; useful when file is confirmed error by GCMS |
What This Means for Applicants Right Now
Until IRCC discloses what generative AI tools it uses, what prompts are issued, and how output is reviewed, applicants should treat the following as practical risk-management steps:
- Request GCMS notes immediately after any refusal. This is the only way to see what the officer actually reviewed. If the notes reference a summary document that does not match your original submissions, that is a potential hallucination flag.
- Cross-check refusal reasons against your actual submissions. If the cited duties, qualifications, or circumstances do not appear anywhere in your file, the error may be AI-generated rather than a judgment call. These cases have a stronger reconsideration or judicial review basis than credibility refusals.
- Observe the JR deadline. For applicants inside Canada, the deadline to file for leave for judicial review is 15 days from the refusal. Reconsideration does not pause this clock. As documented in PGWP Refused for Missing Language Test, missing the JR deadline eliminates what is often the most reliable avenue for correction.
- Document the contradiction explicitly. In any reconsideration letter, identify specifically which duties or facts in the refusal letter do not appear in your application, and which document in your file demonstrates what you actually submitted. Make the officer's task of correcting the error as simple as possible.
- Consider whether your profile warrants a concurrent pathway. While a reconsideration or JR proceeds, assess whether reapplication or a parallel pathway is viable. The PNP Program Finder can identify provincial nominee streams that may provide an alternative route while a federal application error is being resolved.
My Actual Take
When this case came up in conversation among colleagues, the immediate reaction from everyone at the table was the same: the disclaimer is the problem, not just the error. IRCC disclosing AI use in a refusal letter, while simultaneously denying that AI influenced the decision, is not transparency. It is the appearance of transparency that makes accountability harder, not easier.
What concerns me professionally is the verification question. IRCC's AI strategy states that human oversight is central. If human oversight produced a refusal citing electrician duties for a post-doctoral immunologist, the oversight mechanism is not functioning. Either officers are rubber-stamping AI summaries without reading source documents, or the tool is being used at a stage of processing where the officer never sees a full file at all. Neither is acceptable, and neither is disclosed.
The irony is that IRCC faces a genuine processing challenge — approximately one million applications currently exceed service standards. Efficiency tools are not inherently wrong. But generative AI, specifically, is the wrong tool for the task of summarising skilled work experience for NOC classification. It hallucinates. That is a documented, well-understood property of large language models. Deploying a hallucination-prone tool in a process where a single fabricated sentence can end someone's career in Canada is not an efficiency gain — it is a liability that the department will continue to carry until it provides the transparency it has so far refused to offer.
Adé's file was reopened. That is the right outcome. But it should not have required a national newspaper investigation to achieve it. Book a consultation → if you have received a refusal that cites facts that do not match your file.
Frequently Asked Questions
Did IRCC's AI system directly refuse Kémy Adé's application?
IRCC maintains the refusal decision was made by a human officer. However, the refusal letter contains a disclaimer explicitly disclosing the use of generative AI in application processing — and the fabricated job duties that formed the basis of the refusal do not appear anywhere in her submitted documents. Her file was reopened after her lawyer filed a reconsideration request.
What is AI hallucination in the context of immigration applications?
Hallucination refers to a language model generating plausible-sounding but factually incorrect content by drawing patterns from training data rather than accurately reading source documents. In Adé's case, the model appears to have attributed an industrial electrician's job duties to a health scientist, likely by conflating unrelated occupational descriptions during a file summarisation task.
Does IRCC use AI to refuse applications?
IRCC's published AI strategy states it does not use autonomous AI systems to refuse applications, and that all decisions are made by human officers. However, generative AI is being used to support processing — including summarising files — and the Adé case demonstrates that AI-generated content can enter the reasoning of a refusal without adequate verification. As of June 2026, IRCC has not disclosed which specific tools are used or how human verification is conducted.
What should I do if my IRCC refusal letter cites facts that don't match my file?
Request your GCMS notes immediately to see what the officer reviewed. If the cited duties or qualifications do not appear in your original documents, document the specific contradiction and submit a reconsideration request with a clear side-by-side comparison. For applicants inside Canada, file for leave for judicial review within 15 days of the refusal — reconsideration does not pause this deadline. Consulting a licensed RCIC or immigration lawyer before acting is strongly recommended.
What is the 15-day judicial review deadline and why does it matter here?
Under the Federal Courts Act, applicants inside Canada have 15 days from the date of a refusal to file for leave for judicial review at the Federal Court. This deadline is strictly enforced. If an AI-related error is the basis of your refusal, judicial review is often more reliable than reconsideration — but only if the deadline is met. Waiting for reconsideration results while the JR window closes is a common and costly mistake.
Could this type of AI error affect work permit or study permit applications?
IRCC's AI strategy confirms that generative AI tools are being used across multiple application types for summarisation, research, and processing support — not only permanent residence. Work permit and study permit applicants with complex profiles should request GCMS notes and carefully review all cited reasons against their actual submitted documents if they receive an unexpected refusal.