The 2026 Guide
Index

Ethics and bias

Ethics and bias is about building a service that treats people fairly and stays accountable for the decisions it makes, especially when those decisions are automated or use AI. Bias is a design problem before it is a technology problem: any service can advantage some people and shut others out without anyone noticing. When a service automates or assists a decision about a person, the Government of Canada requires it to be transparent, accountable, and fair, and to be assessed for its impact before it goes live. The decisions that shape this are made early and revisited as the service learns.

What good looks like

  • Before launch, the service is checked for who it affects differently and who might be left out, using GBA Plus.

  • If the service automates or assists a decision about a person, an Algorithmic Impact Assessment is done before launch, published, and redone when the system changes.

  • The oversight, testing, and review match the decision's impact level: the higher the impact, the more is required.

  • The data and the outputs are tested for bias, before launch and as the service runs.

  • A person can get a meaningful human alternative to an automated decision, and can challenge or appeal the result.

  • People are told when a decision is automated, and can get a plain explanation of how it was reached.

  • A named person stays accountable for the decisions; the system never is.

  • Any use of generative AI follows the FASTER principles, fair, accountable, secure, transparent, educated, relevant.

Why it matters

Automated decisions can deny a benefit, flag a person, or rank an application, and when the system is biased or unexplained, real people are harmed and the harm repeats at scale. Trust in the service, and the basic fairness owed in any government decision, are on the line. It is also a legal duty: the Directive on Automated Decision-Making makes the impact assessment, notice, explanation, bias testing, human oversight, and recourse mandatory for federal automated decisions, and GBA Plus is required in Treasury Board submissions. One trap worth naming: "representative data" does not fix everything, because systems still misjudge the people who sit far from the average, so testing and human judgment stay necessary.

Whose job it is

Ethics and bias is shared across the team, with each role holding a different part:

  • Data scientists and developers build and test the system, and check the data and the outputs for bias.
  • The department's data, privacy, and legal teams advise from the concept stage, run the peer review, and help with the impact assessment and GBA Plus.
  • The business owner of the application makes sure the assessments are done, that human oversight and recourse exist, and answers for decisions that are fair and lawful.

A closer look

Comparison

Two ways to automate a decision

Pax

Meet Pax, a benefits caseworker. They run a service that screens benefit applications and treated the auto-scoring as a decision about people:

  • ran a GBA Plus and an Algorithmic Impact Assessment before building, and published it
  • kept the AI as a recommendation, with a person making the final call
  • tested the outputs for bias across groups, before launch and as it ran
  • told applicants a tool was used, gave a plain explanation, and offered an appeal to a human

The result: decisions people could understand and challenge, less bias, and a system that holds up to scrutiny.

What Ethics and bias looks like in each phase

The fairness work changes shape across the life of a service.

The fairness work is cheapest and most effective before launch. The team runs a GBA Plus to see who the service affects differently, and if it automates a decision, completes an Algorithmic Impact Assessment to set the impact level and the safeguards. Human oversight, a plain explanation, and a way to appeal are designed in, and the data and outputs are tested for bias before anything goes live.

The official instruments behind ethics and bias

Everything official this subject brings with it, and where in a service's life each one comes up. The full detail, including who does the work and what the business owner personally does, is in the table on the home page.

  • Algorithmic impact assessment (AIA)Only ifsourceAssessment

    A questionnaire the department fills in about itself, scoring how much an automated decision could affect people's rights, health, economic interests or the ongoing sustainability of an ecosystem. The score sets obligations for explanation, human involvement, testing and recourse.

    • AlphaCheck
    • BetaFillSubmit
    • GrowthKeep current
    • MaturityKeep current

Further reading

Beyond the binding Directive, Algorithmic Impact Assessment, and GBA Plus already linked above, a few sources go further. The twelve guiding principles for the use of AI in government set out the openness-and-accountability expectations, and the OPC's principles add the privacy-and-fairness lens. For a plain statement of the ethical values your service should uphold, the made-in-Canada Montréal Declaration on Responsible AI gives ten principles you can hold a design up against, and CIFAR's AI & Society work under the Pan-Canadian AI Strategy gathers Canadian research on AI's effects on people. When you want a hands-on sense of how to check a system rather than just principles, the UK government's Introduction to AI assurance walks through the techniques teams use to test that an AI system is fair and works as intended. For how other places frame it, the US NIST AI Risk Management Framework and the EU AI Act are useful companions.

See also

Assumptions this page makes

You are already working to the Government of Canada Digital Standards, design with users, iterate and improve frequently, work in the open, use open standards, address security and privacy, build in accessibility, empower staff, be good data stewards, design ethical services, and collaborate widely, and to the law on privacy, security, official languages, and accessibility. The standards say how the government works in the digital world. The six Government of Canada digital competencies say what every public servant has to be able to do to work that way, and the team page covers them. This guide builds on those.