Why the courts acted now?
The preamble is candid about the context. Generative AI tools are already being used by litigants and their representatives, and their use is growing. At the same time, none of the four courts has an institutional AI tool that is explicitly authorized, secured and deployed in its own environment.
That gap is the reason for the cautious approach. The guidelines also point to the opacity of how these systems work, and say that opacity alone justifies prudence.
The core premise: AI is a secondary tool, never a source of judgment
The guidelines are built on a clear view of what generative AI is and is not.
According to the courts, it does not understand the law, exercises no judgment, assumes no responsibility, and cannot independently explain its reasoning in a legal sense.
From that premise follow two firm positions:
- Generative AI may only ever serve as a secondary assistance tool. It cannot be a source of authority, legal reasoning, evidentiary assessment or decision.
- No judge may delegate their decision-making function to an AI system, in whole or in part, no matter how sophisticated the system is.
The preamble closes with a single bolded line that sums up the whole document: judging is an exclusively human act.
Accountability stays with the judge
The judge remains fully responsible for the legal reasoning, the content of decisions, the accuracy of facts found, the accuracy of references, citations and authorities, and the consequences of any AI use. The guidelines state plainly that no error can be attributed to the tool. Using AI reduces neither the judge’s duty to verify nor their responsibility for the final result.
What judges should not use it for
- Substantive legal reasoning
- Analysing and characterizing facts
- Assessing evidence and credibility
- Determining the outcome of a case
- Drafting, in whole or in part, the analytical, justificatory or dispositive reasons of a decision
- Using AI to find information or analysis from outside the record that could influence the decision without having been tested in adversarial debate
Section 4.3 then adds a bright-line test. If the purpose or effect of using AI is to determine the legal outcome, assess credibility, interpret contested facts, interpret the relevance of facts, or formulate the reasons for a decision, it must not be used. Notably, this is the one place where the language shifts from “not recommended” to “must not.”
What may be considered, under strict conditions
- Purely linguistic correction, revision, rephrasing or translation of text the judge has already written; expressly excluding draft judgments
- Certain administrative tasks unconnected to judicial reasoning
- Preparing a factual chronology from elements the judge has already identified
- Producing a strictly descriptive summary of testimony, a document or a factual sequence, with no assessment, interpretation, normative selection or conclusion
- Locating information within defined, closed sources (a specific document given to the tool, as opposed to an open web search), subject to systematic verification at the source
The most useful idea in the whole document sits in this section: the distinction between descriptive selection and normative selection. Descriptive selection means restating what was said or written, in what order and by whom, in a way that can be verified against the source. Normative selection means deciding what is relevant, decisive or credible. Only the first may be assisted by a tool. The second belongs to the judge alone.
Every contemplated use is subject to five conditions: full and continuous human control; independent verification of all generated content (back to the judge’s notes, the recording, the transcript, the source document or the original decision); no confidential or identifiable data; an understanding of the tool’s limits; and the ability to explain and fully own the final result. The guidelines’ fallback rule: when in doubt, abstain.
Confidentiality sets the practical ceiling
The obligations of confidentiality and secrecy of deliberations apply in full. Judges should not give an AI tool draft judgments, notes on an ongoing file, or any information protected by law or court order. They must also confirm that the tool’s settings prevent their inputs from being used for training or improving the system, and abstain if they are unsure.
Three further points stand out:
Transcription tools. No transcription or recording-processing tool has been institutionally evaluated, so submitting such content to generative AI is not recommended until one has.
Hidden instructions. Documents from parties or third parties should not be fed to an AI tool as-is. The courts explicitly warn that such documents may contain hidden information or instructions capable of manipulating the tool. In other words, the bench has named prompt injection as a live risk.
Deliberation data. The self-assessment grid observes that, in practice, very little data connected to a deliberation falls outside the protected categories, and that doubtful cases should be treated as protected.
Read together, the permitted uses and the data restrictions pull against each other. A descriptive summary of testimony is on the permitted list, but testimony in an active file will rarely be free of identifiable information. In practice, the room left for AI in the current environment is very narrow.
Competence, anchoring and the self-assessment grid
The guidelines expect judges to develop adequate technological competence: understanding what these tools can and cannot do, recognizing hallucinations, bias, factual errors and cognitive anchoring, and accepting that apparent efficiency gains never remove the need for full verification and may be cancelled out by it. The courts commit to periodic training, up-to-date information on risks and best practices, and ongoing institutional support.
Annex A offers an eight-step self-assessment grid covering the nature of the task, the role of AI, the type of content requested, the absence of interpretation, the data used, verification, cognitive anchoring, and public confidence. The courts are careful to say it is optional, has no normative force, and is not a compliance test.
Its anchoring step is worth highlighting. It asks whether the judge formed their own analysis before consulting the tool, avoided using generated content as a drafting starting point for analysis or reasons, and kept a record of their own prior work. The final step asks whether the use would survive being explained to a colleague, an institutional review, and hypothetical public disclosure.
Recommendations, built to evolve
The guidelines are framed as recommendations, not binding rules, with the exception of the “must not” language in section 4.3. They are also explicitly an evolving framework, to be reviewed periodically in light of technological change, experience within the courts, the possible arrival of secure institutional environments, and best practices in Canada and comparable jurisdictions.
What this means for lawyers and legal tech
For counsel: the bench has set out, in detail, the standard of verification and accountability it applies to itself. It is reasonable to expect no lower standard for AI-assisted material that lands in front of it. The explicit warning about hidden instructions in party documents also signals that courts are thinking about the integrity of what gets filed, not just what gets decided.
For legal technology builders: the guidelines read like a product requirements document for any tool that hopes to serve courts one day:
- No training on user inputs, verifiable by the user
- Retrieval from closed, defined sources rather than the open web
- Output that can be traced back to the exact source passage
- A clear separation between descriptive assistance and anything that ranks, weighs or concludes
- Defences against instructions embedded in uploaded documents
- Institutional evaluation before deployment, especially for transcription
The four chief judges put it simply: judging is and must remain a profoundly human act. Tools that respect that line and can prove it are the ones that will earn a place in the courtroom.