Human in the loop means a person must check or approve an AI system’s output at a defined point before it takes effect — the AI can draft, suggest or decide, but a person is the one who lets the result go forward.
The Federal Court names this exact principle, in these words, twice. Its 2024 notice on AI in court filings states: “‘Human in the loop’: To ensure accuracy and trustworthiness, it is essential to check documents and material generated by AI. The Court urges verification of any AI-created content in these documents”. Its 2025 Interim Principles restate the same idea for the Court’s own use of AI: “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”, and, elsewhere on the same page: “The Court will not use AI, and more specifically automated decision-making tools, in making its judgments and orders, without first engaging in public consultation”.
The Treasury Board’s Directive on Automated Decision-Making builds the same checkpoint into how federal departments must run an automated system — it binds government, not private business, but shows the mechanism in force: notice before a decision is made or assisted by the system, and separately requires: “Providing a meaningful explanation to clients of how and why the decision was made” after it is. Both examples put the human checkpoint at a specific, named moment — not as a vague expectation of “oversight” somewhere in the process.
An agent that drafts a termination letter can be left to draft it end to end; the human-in-the-loop line is drawn at send — a person reads it and clicks send, or the draft never reaches the employee. Move that checkpoint earlier, to reviewing the instructions the agent was given instead of its output, and it stops being the same control.
See also: what are AI guardrails, what is an AI agent, what human in the loop really means.
Deciding where that checkpoint sits, and keeping it there as a system changes, is ongoing ai-operations work, not a one-time design choice.
