Artificial intelligence in public decision-making: Guidelines, jurisprudence and open questions

Like many of us, Canadian public authorities and courts are increasingly using artificial intelligence (“AI”) tools in their day-to-day work. While AI tools may provide welcome administrative efficiencies, their use by decision-makers raises important concerns. The most pressing are (i) ensuring that the decision-makers—and not the AI tools—are grappling with the substance of, and ultimately deciding, the matters before them; and (ii) ensuring that the resulting decisions are procedurally fair to those they affect.

Many Canadian decision-makers are, laudably, working through these concerns in public. This article summarizes the current, public guidelines issued by Canadian courts and governments on the use of AI in administrative and judicial decision-making. It then identifies a number of lessons from the evolving case law on the use of AI and raises some important open questions.

Ultimately, whether AI has affected an individual’s or business’s rights to procedural fairness or produced an unreasonable decision depends on the circumstances. If you are concerned that an administrative decision was unfair or unreasonable, feel free to contact any member of our Administrative and Public Law Group for assistance.

Section 1: Guidelines governing AI in administrative decision-making

Currently, there are not many public rules or guidelines specifically governing the use of AI by administrative decision-makers, tribunals, or courts. We summarize them in this section.

Canadian Judicial Council Guidelines for the Use of Artificial Intelligence in Canadian Courts

On October 24, 2024, the Canadian Judicial Council (“CJC”), a national body that oversees federally appointed judges in Canada, released its Guidelines for the Use of Artificial Intelligence in Canadian Courts (the “CJC Guidelines”).^[1]

The CJC Guidelines expressly deal with the first concern identified above. They emphasize that “no judge is permitted to delegate decision-making authority, whether to a law clerk, administrative assistant, or computer program, regardless of their capabilities.” However, the CJC Guidelines go on to state that judges are permitted to delegate non-judicial decision-making tasks such as proofreading and dictation.^[2]

The CJC Guidelines provide seven guiding principles:

1. Protect judicial independence: The planning, procurement, and deployment of generative AI in Canadian courts must firmly uphold the fundamental principle of judicial independence, encompassing its individual and institutional dimensions.
2. Use AI consistently with core values and ethical rules: Any consideration of the use of assistive AI by judges should always be consistent with the core values of the court^[3] and judicial ethics.^[4]
3. Have regard to the legal aspects of AI use: Any integration of AI into court processes must adhere to all applicable laws. In particular, courts should consider whether generative AI models were trained on data that was acquired unlawfully.
4. AI tools must be subject to stringent information security standards (and output safeguards): Judges must recognize and endeavour to prevent the security and privacy risks associated with using generative AI.
5. Any AI tool used in court applications must be able to provide understandable explanations for their decision-making output: Explainability is the need for such AI tools to provide clear, understandable explanations for their output, making it easier for users (and for those affected) to interpret, trust, contest, or accept AI output in critical workflows. Explainability is akin to the requirement for judges to provide reasoned explanations for their decisions.
6. Courts must regularly track the impact of AI deployments: Courts must do an impact assessment before introducing AI tools and on a continual basis to monitor the development of AI tools and their long-term impacts.
7. Develop a program of education and provide user support: AI should not be employed without users undergoing a comprehensive educational process and understanding best practices for interacting with the technology.

The CJC Guidelines “aim to provide Canadian judges with a principled framework for understanding the extent to which AI tools can be used appropriately to support or enhance the judicial role.”^[5] According to the CJC, the CJC Guidelines are “aspirational and advisory in nature.”^[6] However, a court could adopt a portion of the CJC Guidelines into the common law.

Federal Court Interim Principles and Guidelines on the Court’s Use of Artificial Intelligence

On September 29, 2025, the Federal Court published its Interim Principles and Guidelines on the Court’s Use of Artificial Intelligence (the “Federal AI Guidelines”).^[7]

The Federal AI Guidelines require that the Federal Court:

• not use AI in making its judgments and orders without first engaging in public consultation;
• to consult with relevant stakeholders, if a specific use of AI by the Federal Court will have an impact on the legal profession or the public; and
• embrace the following principles when using AI:
  • Accountability: The Court will be fully accountable to the public for any potential use of AI in its decision-making function.
  • Respect of fundamental rights: The Court will ensure its uses of AI do not undermine judicial independence, access to justice, or fundamental rights, such as the right to a fair hearing before an impartial decision-maker.
  • Non-discrimination: The Court will ensure that its use of AI does not reproduce or aggravate discrimination.
  • Accuracy: For any processing of judicial decisions and data for purely administrative purposes, the Court will use certified or verified sources and data.
  • Transparency: The Court will authorize external audits of any AI-assisted data processing methods that it embraces.
  • Cyber security: The Court will store and manage its data in a secure technological environment that protects the confidentiality, privacy, provenance, and purpose of the data managed.
  • “Human in the loop”: The Court will ensure that members of the Court and their law clerks are aware of the need to verify the results of any AI-generated outputs that they may be inclined to use in their work.

To our knowledge, the Federal Court has not conducted public consultations on the use of AI in making its judgments or orders. If correct, the Federal AI Guidelines currently preclude the Court from using AI “in its decision-making function” (presumably, when deciding the substance of judgments or orders).

Government of Canada’s Directive on Automated Decision-Making

On June 24, 2025, the Government of Canada updated its Directive on Automated Decision-Making (the “Federal Directive”) which applies to nearly all federal departments, boards and tribunals.^[8] The Federal Directive applies to all automated decision systems developed or procured after April 1, 2020, used to make an administrative decision or a related assessment about a client.

The Federal Directive is intended to ensure “that automated decision systems are used in a manner that reduces risks to clients, departments and Canadian society, and leads to more efficient, accurate, consistent and interpretable decisions made pursuant to Canadian law.”^[9]

The Federal Directive requires that applicable federal decision-makers:

• provide notice that a decision will be made or assisted by an automated decision system before making the decision;
• publish a meaningful explanation on how the automated decision-making tool works;
• provide a meaningful explanation to clients of how and why decisions were made; and
• document the decisions and assessments made or assisted by automated decision systems (among other requirements).

All administrative decisions must be reasonable; that is, they must comply with the legal and factual constraints bearing on them.^[10] While government policies are not legally binding, they can assist a court in determining whether an interpretation or a decision is reasonable.^[11] As such, a department’s failure to comply with the Federal Directive could bear on the reasonableness of its decision.

Federal manuals

Under section 5(1)(c) of the federal Access to Information Act, the Government of Canada must annually publish “a description of all manuals used by employees of each government institution in administering or carrying out any of the programs or activities of the government institution.” This includes manuals for any AI tools or systems.

Government of Ontario’s Responsible Use of Artificial Intelligence Directive

On January 7, 2026, the Government of Ontario updated its Responsible Use of Artificial Intelligence Directive (the “Ontario Directive”) which sets out the requirements for the transparent, responsible and accountable use of AI by the Government of Ontario.^[12]

The Ontario Directive sets out six principles:^[13]

1. AI is used to benefit the people of Ontario.
2. AI use is justified and proportionate, and AI systems used are reliable and valid.
3. AI is used in a safe, secure and privacy protective way.
4. AI use is human rights affirming and non-discriminatory.
5. AI use is transparent and meaningful explanations of decisions are made available.
6. AI use is accountable and responsible.

Under the Ontario Directive, if the public is interacting directly with a service that leverages AI (e.g., a chatbot) or if AI is involved in decision-making directly affecting a member of the public (e.g., determining eligibility for a government service or benefit), then the Government of Ontario must:^[14]

1. publicly disclose AI use as part of the process, service, or program; and
2. provide an accessible avenue for the public to seek information about the use of AI in the process, service, or program.

The Ontario Directive states the second requirement does not create a new avenue of seeking review of decisions and reaffirms that pre-existing legal avenues to challenge a decision or outcome of a process, service or program continue to apply. Individuals can continue to challenge Ontario decisions through statutory rights of appeal or applications for judicial review.

Tribunals Ontario’s Practice Direction on the Use of Artificial Intelligence (AI) in Tribunal Proceedings

Section 27 of Ontario’s Statutory Powers Procedure Act requires that the Ontario tribunals to which it applies publish their rules and guidelines, including rules and guidelines governing AI.^[15]

In April 2025, Tribunals Ontario, a group of 12 Ontario tribunals, published their Practice Direction on the Use of Artificial Intelligence (AI) in Tribunal Proceedings (the “Tribunals AI Direction”).^[16] Consistent with concern (i) identified above, the Tribunals AI Direction forbids its Tribunal members from using AI to write their decisions or analyze evidence.

The Tribunals AI Direction allows parties to use AI in proceedings before its Tribunal members; however, it reminds parties that they remain responsible for the accuracy of their written and oral submissions even if they used AI to help prepare them.

Section 2: Key takeaways from the AI and administrative law case law

Any use of AI must be fair

Administrative decision-makers owe individuals in administrative proceedings a duty of procedural fairness.^[17] The level of procedural protections an individual is entitled to depends on all the circumstances, including:^[18]

• the nature of the decision: the more the decision is akin to a judicial decision, the greater the procedural protections owed.
• the nature of the statutory scheme: statutes can outline particular procedures to be followed. Final decisions require greater procedural protections than interim decisions.
• the importance of the decision to the affected person: the more important the decision is, the greater the procedural protections owed.
• the legitimate expectations of the person challenging the decision: if an individual has a legitimate expectation that an administrative decision-maker will follow a particular process, they must follow the particular process. An individual cannot have a legitimate expectation that a particular result or outcome will be reached.
• the agency’s choices about procedures: an administrative decision-maker’s guidelines, policy manuals and handbooks and resource constraints can affect what procedural protections are owed.

AI merely assisting in making a decision does not make the decision procedurally unfair.^[19]

However, early cases suggest that where AI assistance is used, the method(s) by which it was employed should be disclosed to the affected party.

For instance, in Barre v. Canada,^[20] the applicants alleged that the Minister of Citizenship and Immigration relied on racially biased AI photo comparison software to reject their refugee applications. The Federal Court held that the Refugee Protection Division erred in allowing the Minister to admit photo comparisons without disclosing the methodology that was used.^[21]

Similarly, in Ali v. Canada, the Federal Court held that the Refugee Protection Division violated the applicant’s right to procedural fairness when it:^[22]

• “denied his request for further information about the source and methodology used by the Minister in obtaining and comparing the photographs, thereby blocking the Applicant's attempts to test the reliability of the evidence being used against him”; and
• “by accepting without further examination statements by counsel for the Minister that no facial recognition technology was used and the photographs were discovered and compared manually.”

Any use of AI must be reasonable

Generally, the law only requires administrative decision-makers to make reasonable decisions, not ideal or perfect decisions.^[23] To be reasonable, an administrative decision must be:

• justified: the decision is justified against the relevant factual and legal constraints (e.g., the governing statutory scheme);
• intelligible: the person to whom the decision is addressed must be able to understand the decision-maker’s explanation for making the decision;
• transparent: all relevant information on which the explanations rest must be disclosed to the subject of the decisions (with few exceptions).

The Federal Court has held that AI merely assisting in making a decision does not, without more, make the decision unreasonable.^[24]

In a series of cases, the Federal Court has examined the use of Chinook, an AI-assisted tool, used in visa processing.^[25] The Federal Court has consistently held that an immigration officer’s use of Chinook to process visa applications does not, on its own, raise issues of reasonableness or procedural fairness.^[26]

Administrative decision-makers enjoy a presumption that their decisions are lawful and reasonable. The burden is on applicants challenging a decision to prove otherwise.^[27] Therefore, applicants likely need to have clear evidence to prove that a decision-maker’s use of AI was unreasonable or procedurally unfair to set aside a decision on that ground.

In Haghshenas v. Canada, the Federal Court rejected the applicant’s bald allegations that Chinook was not reliable or efficacious.^[28] However, in a subsequent case, Ali v. Canada, the Federal Court held that it was procedurally unfair to simply accept counsel’s statements that its client manually compared photographs instead of using facial recognition technology.^[29]

The Federal Court jurisprudence on Chinook emphasizes that the challenged decisions were reasonable and procedurally fair because a visa officer, and not Chinook, made the final decision.^[30]

Certain AI tools that do not provide a basis for their outputs or otherwise explain their reasoning—so-called “opaque” or “black box” systems—raise specific concerns. If a decision maker relies on the outputs of such a tool, without knowing how those outputs were generated, it is not at all clear how its resulting decision can be transparent, justified, and intelligible.

In Sopeyin v. Canada, the Federal Court held that an immigration officer refusing a work permit based on evidence of “unknown provenance” prevented the Court from assessing the decision and undermined the “the transparency, justification and intelligibility of the Decision and therefore its reasonableness.”^[31] The Federal Court reaffirmed that procedural fairness required the Immigration Officer to give the applicants an opportunity to respond to evidence before relying on it to reject their applications. On the strength of Sopeyin, it is at least arguable that the outputs of any “black box” AI tools, being of “unknown provenance” could also undermine the reasonableness of any resulting decision.

Generally, parties are entitled to use AI to assist them in challenging administrative decisions

Generally, tribunals and courts require applicants to disclose the use of AI tools in their submissions or verify the output of AI tools instead of outright banning the use of AI tools.^[32] The case law has identified some restrictions on parties’ ability to use AI, even where it is disclosed.

AI prompts and outputs have been held not to be “evidence.” In Kevin Huggins v. Red Lobster,^[33] the applicant attempted to enter the questions and answers he received from an AI tool into evidence. The Ontario Labour Relations Board (“Board”) held that these documents were submissions, not evidence.^[34] The Board also held that attempting to introduce these documents was an abuse of process because it was in violation of the Board’s previous directions.

In at least one case, following the filing of written submissions that included fictitious or irrelevant citations, a Court ordered counsel to personally prepare new submissions without using any generative AI for legal research.^[35]

Cautioning parties not to misuse AI is not, without more, procedurally unfair

In Internet Sciences Inc v. CNSX Markets Inc,^[36] the applicant accused a Capital Markets Tribunal adjudicator of being partial for (i) suggesting that the applicant’s references to non-existent or misstated Rules were caused by misusing generative AI and (ii) reminding all the parties that they should ensure the accuracy of their materials regardless of whether they used generative AI.^[37]

The Capital Markets Tribunal adjudicator held that her comments did not indicate that she would not decide the matter fairly. She declined to recuse herself.

Section 3: Open questions

Below, we identify a few questions that the evolving case law has not yet clearly answered.

Do administrative decision-makers need to explain to the public the extent that they are using AI tools to replace human input or human decision-making?

In Haghshenas v. Canada, the Federal Court rejected the applicant’s argument that an administrative decision cannot be reasonable “until it is elaborated to all stakeholders how machine learning has replaced human input and how it affects application outcomes.”^[38]

Subsequently, the Government of Canada introduced the Directive on Automated Decision-Making which requires federal decision-makers to meaningfully explain to the public how automated decision-making tools work. However, it is unclear whether the public has any legal recourse if a federal decision-maker does not follow the Federal Directive. The Government of Canada could internally discipline federal decision-makers who are non-compliant with the Federal Directive.

Are applicants entitled to the AI software used to make decisions that affect them?

The amount of information administrative decision-makers must disclose depends on the importance of the decision. For example, individuals are entitled to more information when facing criminal charges than regulatory charges.

In R. v. Siddiqui, the Court found that facial recognition software and its results are relevant and must be produced to the accused in a robbery and assault trial.^[39]

In the R. v Hughes companion cases, the Court found that an automated software used to collect evidence against the accused was relevant, but the public’s interest in protecting police investigatory techniques outweighed the accused’s legitimate interest in disclosure of operational copies of the software.^[40] The Court also found that the accused was not entitled to the source code for the software, or the training materials for the software because they were not relevant to the criminal charges.^[41] The Court ordered the Crown to produce the software’s manuals, logs, transcripts, and validation test reports after permitting the Crown to make redactions for lack of relevance and privilege.^[42]

In May v. Ferndale Institution,^[43] the Supreme Court of Canada held that procedural fairness required that Correctional Service of Canada disclose a computerized scoring matrix to inmates before relying on it to make decisions on transferring inmates.

In Re Lamarche,^[44] the accused, who was facing regulatory securities charges, brought a preliminary application requesting the disclosure of technical information relating to the software the Director used to collect evidence against him (among other relief sought). The British Columbia Securities Commission held that the architecture of the e-discovery software and the software provider’s security protocols was not relevant and need not be disclosed to the accused.^[45]

How much can AI influence the decision-making process before it becomes the decision-maker?

It would be unreasonable for a decision-maker to accept, wholesale, the outcome of an AI tool, or to adopt the reasons of an AI tool without understanding how the AI reached that outcome. Administrative decision-makers are prohibited from fettering their discretion, and cannot abdicate their responsibility to decide matters delegated to them by Parliament or the legislature. However, the Federal Court suggests in Pjetracaj v. Canada, that an applicant needs “clear evidence” to prove that an administrative decision-maker inappropriately delegated its decision-making power to an AI tool.^[46]

Provided an administrative decision-maker actually makes the decision, the common law likely does not prohibit them from using AI to help draft their reasons. In Boukhanfra v. Canada (Minister of Citizenship and Immigration), the Federal Court held that decision-makers may make use of boiler-plate statements or reasons, but cautioned that boiler-plate language alone does not render a decision immune from judicial review.^[47]

Are applicants entitled to have a human make administrative decisions that affect them?

In Haghshenas v. Canada, the Federal Court held that the fact that a visa decision was reached with the assistance of AI was not relevant to the duty of procedural fairness.^[48]

However, it is far from clear that the use of AI is never relevant to the duty of procedural fairness. For example, if an applicant has a legitimate expectation that an administrative decision-maker will follow a particular process (e.g., a human will make a decision without relying on AI), procedural fairness requires that process be followed. Importantly, such expectations extend only to procedure, not to substantive outcomes.