Part Deux - Where the Line Gets Harder to Draw

In Part One, we looked at Québec’s new GenAI guidelines for judges alongside guidance we have helped develop. The approaches begin from the same principle that judgment must remain human. The differences appear when that principle is applied to actual chambers work and the line between assistance and judicial reasoning has to be drawn.

Summarization is a good example. Québec allows strictly descriptive summaries but draws a line against what it calls “normative selection,” meaning decisions about what is relevant, determinative, or credible. That distinction is important because judges should not ask a machine to decide which testimony deserves to be believed or which facts should control the outcome of a case.

But summarization itself requires selection. A model cannot reduce an 80-page transcript to 5 pages without deciding what to include, what to omit, what to compress, and what to emphasize. Those choices do not necessarily amount to judicial reasoning, but neither are they simple reproduction.

Our new Judicious Judge’s Guide to Generative Artificial Intelligence and Large Language Models recognizes that problem directly. It warns that a summary can focus on the wrong issue, take material out of context, or omit something important even when every statement it contains is accurate. It also cautions that a model may identify something as important simply because it occupies more space in the source material.

That is a different problem from the classic hallucination. A summary can contain no invented facts and still give the judge a distorted picture because of what was emphasized or left out. Our response is to treat summarization as useful but bounded assistance, with the judge retaining responsibility for credibility, importance, and legal significance and verifying the output against the underlying source.

The distinction from Part One matters here. We do not regard every act of selection by a GenAI system as judicial reasoning once the judge has independently reviewed the record and developed an understanding of the case. Québec identifies the same risk but draws a tighter line around the kind of selection the system should be allowed to perform.

Draft opinions expose the difference even more clearly. Québec advises against using GenAI to draft analytical, justificatory, or operative portions of a decision and excludes draft judgments from otherwise permissible linguistic revision. It also cautions against providing draft judgments to available GenAI systems because of confidentiality and the secrecy of deliberations.

The guidance we provided in Navigating AI in the Judiciary: New Guidelines for Judges and Their Chambers and Judicious Judge’s Guide to Generative Artificial Intelligence and Large Language Models takes a different approach. Both contemplate editing and proofreading draft opinions under appropriate safeguards while requiring the judge to retain control over substance, tone, reasoning, and the final product.

Here again, we separate the judicial work from the assistance being requested. Once a judge has reviewed the record, reached the necessary preliminary determinations, developed the analysis, and written a paragraph explaining it, asking an appropriately secure GenAI system to improve the clarity of that paragraph, in our view, does not necessarily transfer judicial reasoning to the machine. Nor does asking it to identify an incorrect name, inconsistent date, citation problem, grammatical error, or confusing sentence.

That is different from asking GenAI to supply the reasoning or determine the result. Our Judicious Judge’s Guide generally recommends that judges and clerks prepare the initial draft themselves, in part to reduce anchoring and automation bias. The human should do the judging before the machine assists with the work that follows.

Legal research and bench memoranda reveal the same distinction. Québec allows GenAI to locate information within specified closed sources, subject to verification at the source, but advises against using it for substantive legal reasoning. Our guidance gives legal-specific systems a broader supporting role, including identifying potentially relevant authorities, surfacing omissions in the parties’ research, and helping organize a neutral bench memorandum, assuming proper verification follows.

Our Judicious Judge’s Guide also contemplates using GenAI to identify possible vulnerabilities or analytical gaps in reasoning the judge has already developed. None of those activities removes the judge’s obligation to independently evaluate the law, determine its application to the case, and own the final analysis and outcome.

The line for us is not crossed merely because the tool is contributing to analytical work. The concern is whether GenAI has moved from assisting the judge to supplying the judgment the judge is responsible for exercising. Québec draws that line earlier in several of these settings.

Security presents a related but separate issue. Québec notes that its courts do not currently have GenAI systems that have been institutionally authorized, secured, and deployed for judicial use. That context understandably contributes to its caution about draft judgments, case notes, and protected material.

But the substantive difference between the frameworks would not necessarily disappear if Québec adopted an approved system. A court could evaluate an appropriately configured enterprise or legal-specific system such as Westlaw CoCounsel, Lexis+ with Protégé, or another vetted platform before developing a court-specific tool. Better privacy and security can address whether sensitive material should be placed into a system, but they do not resolve the separate question of what the system should be permitted to do with it.

Those are two different lines. One concerns whether the tool and its environment are appropriate for confidential or deliberative information. The other concerns whether the requested activity remains assistance or has crossed the line into judicial reasoning.

A secure system could still be asked to perform an inappropriate judicial function. Conversely, a task that remains appropriate assistance may still be off limits if the available system cannot adequately protect the information required to perform it. That distinction will become more important as courts gain access to more capable and more secure systems.

Québec itself recognizes that its framework will evolve as technology develops, courts gain experience, and secure institutional environments become available. We agree with that approach because neither framework should be treated as the last word.

There is also much in Québec’s guidance that should endure regardless of where these operational lines ultimately settle. Judges should develop their own analysis before consulting GenAI, verify generated material, remain personally responsible for the final result, and consider how their use of the technology could affect public confidence.

Those concerns bring the approaches back together. Québec has made an important contribution by forcing courts to confront these distinctions directly rather than reducing the issue to whether judges should or should not use GenAI. The more useful comparison is where assistance ends and judicial reasoning begins, and which activities properly fall on either side of that line.

The technology will evolve, and some of those boundaries may move with it. But human judgment must always remain paramount.