Where Assistance Ends and Judgment Begins

For the first time on this Substack, I have a co-contributor. Given the subject, Maura R. Grossman was the obvious person to ask.

Maura is a lawyer, professor at the University of Waterloo in Canada, and technologist who has spent years studying the responsible use of artificial intelligence in law and the courts. We were also among the co-authors of the 2025 Sedona guidance on judicial AI, entitled Navigating AI in the Judiciary: New Guidelines for Judges and Their Chambers, and our new Judicious Judge’s Guide to Generative Artificial Intelligence and Large Language Models. So when Québec’s courts recently released their own GenAI guidelines for judges, it seemed only fitting that we examine them together.

Québec’s courts deserve credit for issuing the guidelines. They have addressed a rapidly evolving technology before either the technology or judicial practice surrounding it has settled and produced a thoughtful framework grounded in judicial independence, accountability, confidentiality, competence, and public confidence.

The guidelines begin in the right place. Judging remains a fundamentally human function. GenAI cannot replace judicial reasoning, the assessment of evidence, or deliberation, and judges remain responsible for the decisions issued in their names.

That principle closely aligns with the guidance we have helped develop over the past two years. The Sedona guidance recognized that GenAI may assist with judicial work but warned against allowing it to undermine human judgment, while our new Judicious Judge’s Guide makes clear that GenAI does not understand law, exercise judgment, or accept responsibility. The tool must remain subordinate to the judge.

The Québec guidelines also counsel caution. They do not encourage judges to use GenAI simply because it is available, they recognize that abstention may sometimes be appropriate, and they anticipate that the framework will evolve with the technology and judicial experience. We share that sentiment. Courts do not need to win a race to adopt artificial intelligence. Their responsibility is to use it carefully and deliberately, and to understand enough about the technology to distinguish useful assistance from unacceptable risk.

There is substantial common ground. Both approaches reject delegating substantive legal reasoning, credibility determinations, the weighing of evidence, or the outcome of a case to GenAI. Both seek to preserve the same principle that judicial authority belongs to human beings who are accountable for exercising it.

But the approaches begin to diverge over where the line between assistance and judicial reasoning should be drawn. Québec places tighter limits on GenAI assistance as a task approaches judicial reasoning or the expression of that reasoning. Our guidance permits more bounded assistance after the judge has independently performed the underlying judicial function and remains responsible for the result.

Our guidance does not treat GenAI assistance as the same thing as judicial reasoning once the judge has reviewed the record, considered the arguments, made the necessary preliminary determinations, and developed an independent analysis. Summarizing part of a record, helping improve the wording of a judge-written paragraph, locating additional authority, or identifying a possible weakness in reasoning the judge has already developed can remain assistance rather than delegation, provided the judge retains control and verifies the work.

That is a difference in where the line is drawn, not a disagreement about the need for a line. Québec is more restrictive as GenAI approaches the reasoning and written expression of a decision. Our guidance permits more assistance near that boundary, but only after the human has gone first and remains accountable for the final product.

The distinction also does not disappear simply because a court adopts a better tool. A court might approve an appropriately configured enterprise or legal-specific system such as Westlaw CoCounsel, Lexis+ with Protégé, or another vetted platform before developing anything bespoke for the judiciary. Better privacy and security may determine whether sensitive information can safely be placed into the system, but they do not answer the separate question of what the system should be allowed to do with that information.

Both approaches also care about sequencing. Québec’s optional checklist asks judges to formulate their own analysis before consulting GenAI and to avoid allowing AI-generated material to become the starting point for their reasons. Our Judicious Judge’s Guide similarly warns about anchoring and automation bias and generally recommends that the human begin the analytical and drafting work before turning to GenAI for assistance.

That safeguard addresses a risk that is easy to underestimate. A judge who first encounters a case through an AI-generated summary may begin with the system’s framing of the dispute, including what it selected, omitted, emphasized, or compressed. Even an accurate and fluent output can shape the analysis before the judge has independently developed one.

Québec and the guidance we have helped develop therefore share a theory of judicial responsibility while drawing some operational lines in different places. Neither approach should be treated as fixed. The technology is changing rapidly, and judicial experience with it is still developing.

What matters is that courts are moving beyond the question of whether judges will encounter GenAI. They are beginning the harder work of deciding how it may assist without displacing the human responsibility at the center of judging. Human judgment must remain paramount, even as thoughtful courts draw different lines around the technology that surrounds it.

In Part Two, we will examine where those differences become most visible in practice, including summaries, legal research, draft opinions, and the security of the systems judges use.