When governments use AI, who is accountable?

You apply for a visitor visa, a benefit or a permit. Weeks later, a decision arrives.

Was AI involved? You may not know.

Government systems can sort applications, flag problems and assess eligibility. Some even make decisions without a person handling every case.

If a system influences the outcome, who is accountable?

The government remains responsible. But who makes the decision, and which safeguards apply, depend on how the system is used. That difference can matter when you need an explanation or want to challenge the result.

Here's the version that fits in your head.

Some tools translate documents, search records or answer routine questions. Others assess eligibility, flag possible errors or decide which applications need closer review.

In November 2025, the federal government launched its first public AI register, listing more than 400 systems across 42 institutions. It included tools already in use and others still being explored or developed.

A register entry doesn't mean a system is making decisions about Canadians. A chatbot that answers routine questions and a system that assesses someone's eligibility raise different questions. What matters is what the tool actually does, not whether its description includes the word AI.

Immigration, Refugees and Citizenship Canada (IRCC) shows why that distinction matters.

IRCC uses advanced analytics to process certain applications. For routine visitor-record applications, a tool can make a positive eligibility determination before an officer checks admissibility and makes the final decision. The officer can overturn the tool's finding.

IRCC says its tools don't refuse applications or recommend refusals. It also says some types of applications can receive fully automated approvals based on rules developed by IRCC officers.

So human involvement isn't identical across every process. And even when an officer decides, software can shape which information reaches them, how a case is categorized and how a file is handled. Human review doesn't automatically catch every error.

Where AI sits in a decision matters as much as whether it's used at all.

Canada has no single set of rules for all government AI. Requirements depend on the institution, jurisdiction and how the system is used.

The federal Directive on Automated Decision-Making, in effect since 2019, applies to covered federal departments when automated systems make or help make administrative decisions about people.

Before deploying a system covered by the directive, departments must complete an Algorithmic Impact Assessment. They must also publish its results. The assessment assigns one of four levels, from Level I (little to no impact) to Level IV (very high impact).

The safeguards increase with impact, reflecting how much a wrong decision could affect someone. At Levels I and II, systems may decide without direct human involvement, with human quality assurance and intervention where appropriate. At Levels III and IV, a person must make the final decision and review system outputs for accuracy and appropriateness. These requirements appear in Appendix C of the directive.

The directive also sets requirements for notice, explanations and opportunities to challenge decisions. It doesn't cover every federal institution or every AI use: drafting or policy research, for example, may fall outside its scope. Other privacy and security obligations can still apply.

The federal directive doesn't automatically apply to provincial or municipal governments.

Ontario's Responsible Use of Artificial Intelligence Directive, effective December 1, 2024, sets AI risk-management, transparency and accountability requirements for provincial ministries and agencies. When AI is involved in decisions that directly affect people, or when people interact with an AI service such as a chatbot, it requires disclosure and a way to ask for information.

Ontario's Enhancing Digital Security and Trust Act, 2024 creates a framework for additional AI requirements in the broader public sector, including risk management and human oversight.

Unlike the directive, the Act's AI-specific obligations depend on regulations.

Governments must follow applicable rules on procedural fairness, privacy and lawful decision-making. Depending on the decision, people may be entitled to reasons, a review, an appeal or judicial review. AI doesn't erase those protections, but they vary by service and jurisdiction.

Responsibility runs through several levels:

These roles are different. A court may examine a particular decision, while an auditor reviews how a government program is run.

Buying an AI system from a private company doesn't transfer public responsibility to the vendor. Governments can contract out technology, not accountability for exercising public authority.

Rules don't answer every question.

Some uses fall outside decision-making directives. Research, drafting and policy analysis can shape government work without being covered by rules for individual administrative decisions.

Disclosure isn't the same as explanation. A public register might identify a system without showing how it affected someone's file.

Human oversight isn't foolproof. People may give an automated recommendation too much weight, even when other information suggests caution. Researchers call this automation bias.

Coverage and information vary. Requirements differ across governments, and registers or assessments may be incomplete or out of date.

None of this proves that an automated decision is unfair. It shows why accountability needs more than saying a human was involved.

You don't need to understand an AI model to ask how a government decision was made.

If you know the name of the system, look for public documentation before assuming how it was used in your case. For federal systems, the AI register and published Algorithmic Impact Assessments are useful starting points.

You may not receive every technical detail, and your options depend on the decision and applicable law. But asking puts the question where it belongs: with the responsible institution.

The Weekly Brief — Issue 12 — The Bank of Canada's forecasting model supports analysis; people remain responsible for policy decisions.

AI agents and agentic AI, explained — What changes when AI moves from answering questions to taking action.

How to check AI answers and sources — Why important AI-generated claims need verification.