The denial letter arrives in an envelope like any other government correspondence. It tells a person with a disability that their claim for benefits has been reviewed and the decision has been made. It does not tell them that the review was partly or fully conducted by an algorithm. It does not name the parameters the system used. It does not explain whether the decision might have been affected by data errors, by patterns the model learned from historical claims that may themselves reflect old administrative biases, or by the kind of confidently wrong output that AI practitioners call hallucinations — fabricated information generated with no basis in fact but presented as a legitimate finding. The person receiving the letter is simply told the outcome. In Quebec, the law now says they have a right to more than that. But extracting that information from the organizations making these decisions remains, in practice, an opaque and contested process.
Federal and provincial authorities are now investigating how AI tools are influencing disability benefit decisions in Quebec, amid growing evidence that automated systems used to process claims are generating errors that claimants cannot identify and challenge. The investigation involves Quebec’s Ministry of Cybersecurity and Digital Technology as well as federal oversight bodies looking at whether current practices comply with privacy and human rights obligations. The specific contours of the investigation have not been fully disclosed, but the concern pattern is consistent with what legal experts have been raising for years: when a benefits denial is based on an AI-generated error or a fabricated claim element, the claimant’s ability to challenge it is fundamentally compromised if they don’t know the AI was involved in the first place.
| Category | Details |
|---|---|
| Subject of Investigation | Use of AI and automated decision-making tools in processing and denying disability benefit claims in Quebec; investigation involves federal and provincial authorities including Quebec’s Ministry of Cybersecurity and Digital Technology (MCN) |
| Quebec Privacy Law | Law 25 (fully in force September 2024) — requires organizations to notify individuals when a decision is made solely by automated processing; individuals have the right to know the factors, information, and parameters used |
| Federal AI Legislation Status | Bill C-27 (including the Artificial Intelligence and Data Act, AIDA) — died on the Order Paper in January 2025; Canada currently has no binding AI-specific federal law; Canadian Human Rights Act is the primary instrument available |
| Core Concern: AI Hallucinations | Automated systems may generate errors or fabricated justifications (“hallucinations”) in denial decisions; if a denial is based on AI-generated errors, legal experts say it could be challenged as bad faith — but only if the claimant can discover the AI was involved |
| Systemic Bias Risk | AI trained on historical disability claim data may encode past administrative biases, disproportionately disadvantaging claimants from certain communities or with certain types of conditions — errors are harder to detect and challenge than in human decisions |
| National Sprint Report | February 3, 2026 — federal government released results of its 30-day “national sprint” AI consultation; report acknowledged privacy, safety, transparency, systemic bias, and accountability concerns raised by over 160 civil society organizations |
| Centre for Policy Alternatives Position | Canada lacks robust, binding legislation for AI in public services — leaving individuals receiving government decisions by algorithm with limited recourse and no mandatory disclosure rights at the federal level |
| Quebec CAI Enforcement | Commission d’accès à l’information (CAI) has broad enforcement powers including audits, binding orders, and administrative monetary penalties for Law 25 violations involving automated decision-making |
Quebec has more legal protection in this area than any other province in Canada. Law 25, the province’s overhauled privacy legislation that came into full force in September 2024, requires organizations to notify individuals when a decision affecting them is made solely through automated processing. Upon request, those individuals have a right to know the personal information used, the factors and parameters that drove the decision, and the opportunity to have a human review it. That framework, heavily influenced by Europe’s GDPR, has been cited by legal commentators as a potential template for the rest of the country. It has also exposed, through its enforcement, how many organizations — including public bodies — were making automated decisions about Canadians while providing no disclosure at all, because nothing required them to.
The federal picture is considerably bleaker. Bill C-27, which included Canada’s first AI-specific legislation under the Artificial Intelligence and Data Act, died on the Order Paper in January 2025 when Parliament was prorogued. No replacement legislation has been introduced. In the absence of binding federal AI law, the primary instruments available to challenge discriminatory or erroneous automated decisions are the Canadian Human Rights Act and provincial human rights codes — neither of which was designed with algorithmic decision-making in mind. The Centre for Policy Alternatives, in its February 2026 assessment, was direct: Canada lacks the robust, binding legislation that would give people receiving AI-generated government decisions any meaningful right to accountability. The federal government’s “national sprint” AI consultation — a 30-day exercise that ended in February 2026 and was itself criticized for using AI to assess the public responses — acknowledged the concerns without committing to a timeline for legislative action.
The problem of AI hallucinations in administrative decisions deserves particular attention because it changes the nature of the legal challenge in ways that haven’t been fully worked through the Canadian court system. In October 2025, a Quebec Superior Court fined a self-represented litigant $5,000 after he submitted court filings containing fabricated citations generated by AI — a case that illustrated how confidently AI systems can produce false information. When that same dynamic occurs inside a benefits adjudication system, the claimant faces an asymmetric situation: the organization made a decision using information the claimant cannot see, and the claimant must challenge it without knowing whether the decision was based on accurate data, a biased pattern from historical claims, or an outright fabrication the system presented as fact.
There is a sense, watching these investigations unfold in 2026, that the gap between how quickly AI has been deployed in public services and how slowly the accountability framework has developed is not an oversight but a choice — one made repeatedly, at multiple levels of government, under pressure from the same efficiency arguments that made the tools attractive in the first place. The disability claimant sitting with a denial letter in hand, trying to understand what happened and whether it was right, is at the far end of that gap. The investigation is trying to close it. It is still unclear whether it can move faster than the technology being investigated.
