Dmitry Shniger says the joint framework catalogues prohibitions without telling judges how to properly use the tech
Quebec's new guidelines on the judicial use of generative artificial intelligence are written to discourage judges from using the technology at all, says Dmitry Shniger, a Toronto lawyer whose research focuses on applying legal design in the public sector to make law and policy more human-centred and accessible.
Chief Justice of Quebec Geneviève Cotnam, Superior Court Chief Justice Marie-Anne Paquette, Court of Québec Chief Judge Henri Richard, and Chief Municipal Judge Nathalie Duchesne are the signatories to the document, released on September 4.
The guidelines come months after La Presse reported that a Quebec judgment contained references to non-existent case law, raising the possibility that generative AI was used to prepare it. A Canadian Lawyer survey conducted last April found the Quebec Court of Appeal and the Quebec Superior Court to be among the courts with the most developed AI policies in the country.
The guidelines' core principle is that judging is an exclusively human act whose decision-making function may not be delegated to AI.
Shniger does not quarrel with that.
“I agree with the substance, but I'm very disappointed with the form,” he says.
As Shniger sees it, effective guidance should tell a user both how to use something and how not to use it.
“The Quebec document is a list of prohibitions and risks, and it leaves a judge with only two options,” he says. “Abstaining whenever the position is unclear – which the guidelines expressly recommend – or concluding that using AI in compliance with them is not possible at all.”
Shniger doesn't object to the provisions that tell judges what they should not use AI for: legal reasoning, analyzing facts, weighing evidence and credibility, deciding outcomes, and drafting any part of their reasons.
His difficulty is with the guidelines' enunciation of what is allowed: a judge may use AI to produce a strictly descriptive summary that interprets nothing and concludes nothing. To paraphrase the guidelines, AI may restate what was said, in what order, by whom, all of which can be verified against the source. Deciding what is relevant, determinative, or credible is the judge's province alone.
As Shniger sees it, the distinction is unworkable.
“Confining AI use to a summary that reproduces what was said while making no judgment about what matters ignores AI's strength, which is pointing out inconsistencies.”
While the guidelines don’t explicitly prohibit judges from feeding confidential data into AI, Shniger says they will have that effect.
“But that's unworkable as well, and the guidelines recognize it,” he says. “They say explicitly that in practice very little material relating to judges' decisions is not confidential, and that anything whose confidentiality is in doubt should be excluded.”
Compounding the problem is the prohibition on feeding documents from parties or third parties to AI, as they may contain hidden instructions that skew the output, unless their content has been appropriately verified. The guidelines, however, say nothing about what constitutes such verification.
And finally, judges may not feed their notes on an ongoing file to AI.
“So what can a judge act on?” Shniger asks. “Take away notes, documents, and confidential information, and what remains? In the end, the guidelines are confusing about anything except corrections and mechanical mistakes.”
The guidelines do explain why the courts drew the lines where they did. No authorized, secure AI tool has been deployed in any of the four courts, the preamble notes, and the commercial tools available run on algorithms whose workings remain largely opaque.
They also commit the courts to periodic training and institutional support – as an accompaniment to the document, not as a precondition to using the tools.
The Canadian Judicial Council issued its own guidelines for the use of AI in Canadian courts in 2024.
“They're quite general,” Shniger says. “The Quebec guidelines are more precise and better written.”
But he says both share a foundational fault.
“They fail to recognize that AI is a disruption to which the courts must adapt because AI isn't merely a tool that works like other existing tools,” Shniger says. “It changes how judges work.”
That approach contrasts with the AI guidance offered by the Action Committee on Modernizing Court Operations, a national body co-chaired by the Chief Justice of Canada and the federal justice minister.
“The committee’s guidance starts with the benefits of AI and how it can improve the judicial system, and it gives judges assistance in deciding whether and how to use it,” Shniger says.
From a broader perspective, Shniger says the Quebec and CJC guidelines identify the wrong risk. The danger is not the technology but users who do not understand how it works – users with unreasonable expectations or users who rely on output because it reads like the real thing.
Judges, Shniger points out, are not in these categories. They are a small, sophisticated, heavily screened group, and it would be much more useful to instruct them on how to use AI rather than merely providing a list of possible misuses.
“Mandatory training for judges before they use the tool would have been a preferable approach than a public document that signals the profession not to use AI.”
The Chief Justice of Quebec was unable to comment by Canadian Lawyer's deadline.
