When AI starts writing awards: a (Quebec) wake-up call?
There is a particular irony in an arbitral award being set aside not because the legal conclusion was necessarily outrageous, but because the reasoning behind it appears to have been built on authorities that never existed in the first place.
In ARIHQ v. Santé Québec, the Quebec Superior Court annulled an arbitral award after discovering that the doctrinal and case law references relied upon by the arbitrator were hallucinations generated by AI.
The underlying dispute was relatively ordinary: a healthcare remuneration disagreement involving procedural deadlines and contractual limitations. The arbitrator dismissed the claim as time-barred. On the surface, there was nothing particularly extraordinary about the outcome.
The problem emerged when lawyers could not verify the authorities cited in the award.
One doctrinal article did not exist. Several cases cited as central authorities either referred to entirely unrelated judgments or were completely fictional. One arbitral authority cited in support of the reasoning turned out to be imaginary altogether.
Courts are generally reluctant to interfere with arbitral decisions merely because an arbitrator may have made an error of law. Arbitration survives on finality. But this case moved beyond legal error and into something more fundamental: whether the arbitrator had effectively delegated the judicial function itself.
The Quebec court drew a line that is not surprising at all: AI may assist decision-making, but it cannot replace the human exercise of judgment.
This case is yet another example of uncritical reliance on AI.
There is now a dangerous tendency in professional environments to confuse polished language with reliability. Large language models produce outputs with extraordinary fluency. The formatting looks convincing. The citations appear legitimate. The tone sounds authoritative. The structure resembles competent legal analysis.
All that glitters is not gold.
I am curious what the prompt looked like.
Not because prompts are inherently problematic, but because the quality of legal AI output depends almost entirely on the quality of human supervision surrounding it. There is a significant difference between:
“Summarise the principles from existing Quebec authorities on contractual limitation periods and provide sources for manual verification”
and:
“Draft an arbitral award rejecting the claim.”
One uses AI as an assistant. The other risks using it as a substitute. Naturally, this statement is oversimplistic, but let’s not undermine the importance of prompting, which is why prompt engineering is flourishing as a discipline now.
There is also something slightly ironic about this occurring in arbitration. Arbitration is often marketed on the basis of expertise, discretion, precision and trust in the appointed decision-maker.
If parties begin to suspect that awards are being generated through unchecked AI processes, confidence in the arbitral system itself becomes vulnerable.
Why does this keep happening?
Because AI is exceptionally good at reducing friction. It accelerates drafting. It shortens research time. It creates the illusion of efficiency. Under pressure, busy professionals are tempted to trust outputs that “look right”.
But legal work is not simply about producing text.
It is about judgment. Verification. Weighting competing authorities. Understanding nuance. Identifying what feels wrong. Recognising when a citation does not fit naturally within the jurisprudential framework. Those instincts are developed through training and experience, and they cannot simply be outsourced to a probabilistic model.
We must get better at prompting while still preserving the core skills that make good lawyers: critical thinking, attention to detail, lateral analysis and the ability to verify large volumes of material under pressure and within tight timeframes. AI can support and enhance those skills, but it cannot replace the professional judgment behind them.
Read more about the decision in ARIHQ v. Santé Québec here.
AN
