Delegatus Non Potest Delegare: Why this Latin expression may mean a lot for AI use by tribunal adjudicators and judges

As AI has reached a stage of ubiquity, some authors have heralded it as a transformative form of technology akin to electricity and the invention of the steam engine.^[1]

In the administrative law world, however, the use of AI by adjudicators, tribunals and judges has serious implications. Because most large language models ("LLMs") obtain their data sets from the Internet, they are capable of both significant errors (what we euphemistically refer to as "hallucinations") and bias. This reality poses a significant challenge for adjudicators who rely on LLMs to assist with their administrative burden and, in some case, draft decisions affecting the rights and privileges of individual claimants.

But the law is not without recourse: traditional administrative legal doctrines can provide a series of checks and balances on the use of AI in the adjudicative process.

This arsenal includes the duty of procedural fairness (such as the doctrine of legitimate expectations), and production requirements where AI is employed in the decision-making process.^[2]

A recent decision of the Quebec Superior Court, AQHIQ c. Santé Québéc,^[3] sheds light on how a rare principle of administrative law, delegatus non potest delegare (hereinafter the "Delegatus Doctrine"), could be used to curtail the improper use of AI by administrative adjudicators.

"A Delegate Cannot Delegate"

In Canadian administrative law, the Delegatus Doctrine creates a general presumption against sub-delegation by a delegate authority.^[4]

The doctrine does not apply, however, "when the action is purely administrative or of such a character that no significant degree of discretion or independent judgment is involved. It only applies to discretionary decisions, legislative or adjudicative decisions".^[5] In other words, a delegate may subdelegate non-decisions to another party. But the moment there is a delegation of discretionary authority, the presumption is invoked.

The presumption, however, can be rebutted where that sub-delegation of authority is justified by "either express or implied statutory authority".^[6]

Even if, however, a delegate has been authorized by statute to sub-delegate that decision-making function, the Court still reserves the right to determine if such delegation was "proper".^[7]

A delegated power is "proper" where it provides for "standards, rules and conditions to guide the decision-making process".^[8] There must be "decisional criteria" that restrain the subdelegate. A subdelegate cannot have unlimited decision-making authority and must be constrained by standards and rules.^[9]

If no such guiding criteria are provided, the delegation of authority is improper.

Delegating the Adjudicative Function to AI

The ARHIQ decision involved a financial claim of approximately $1.2 million by a health care provider against Quebec's Ministry of Health for residential placements.

The claimants pursued their claim at arbitration. Ultimately, the Arbitrator rejected the claimant's dispute as being statute-barred under the applicable limitations legislation.

Upon reviewing the Arbitrator's decision, the lawyers for the applicants discovered that the arbitrator's reasons included articles that did not exist, and four decisions that supported his decision that were also non-existent. As it turns out, the Arbitrator had relied on AI to draft their reasons for decision. As it also turns out, the AI tool used clearly "hallucinated".

The Quebec Superior Court set aside the arbitral award, despite the difficulty in seeking to overturn arbitral awards in the Courts under Quebec's Code of Civil Procedure.

In reaching this conclusion, the Court relied on the Delegatus Doctrine. While the Court took no issue with the use of AI to conduct research or assist with administrative tasks, the Court made it clear that the adjudicator could not delegate the function of drafting a decision to an algorithm or any AI device.

Parties are entitled to expect that a human adjudicator makes the decision affecting their rights and privileges. They have a right to expect that reasons for decision are drafted by a human. They also have a right to expect that an adjudicator will not disclose highly confidential information, particularly in the case of a non-public arbitration, to an AI bot.

Relying on the Delegatus Doctrine, the Court concluded that public confidence in the rule of law and arbitration processes generally means that decision-makers cannot subdelegate the adjudicative function to others, including AI:

La primauté que l'on accorde à l'autonomie de la volonté des parties dans le choix de l'arbitre, l'importance de la rédaction des motifs pour assurer une décision éclairée, le devoir de l'arbitre de maintenir le secret du délibéré et l'impératif de conserver la confiance du public dans le processus d'arbitrage justifient que les décisions soient rédigées par l'arbitre choisi par les parties sans délégation à des tiers.^[10]

A Human Decision-Maker

AQIHQ sets clear lines in the sand regarding the use of AI by administrative decision-makers, as well judges and other adjudicators.

The Delegatus Doctrine permits the delegation of administrative, non-adjudicative tasks to AI or an algorithm. Research, clerical work and even citations arguably fall within this ambit.

However, decisions involving the rights and privileges of the parties, and the reasons that encapsulate them, demand human discretion. The Delegatus Doctrine prevents the very human exercise of decision-making from being delegated to non-human authorities.

As AQIHQ makes clear, any such delegation of adjudicative authority risks undermining the rule of law and tainting decisions with bias and inaccuracies.

Marco P. Falco is a partner in the Litigation Department at Torkin Manes who focuses on judicial review applications and civil appeals. You may contact Marco about your matter at [email protected]. Please note that a conflict search must be conducted before your matter can be discussed.