Haghshenas v Canada (Citizenship and Immigration), 2023 FC 464 (March 31, 2023)
Category: Reasonableness review, procedural fairness, AI decision-making.
Context: The applicant sought judicial review of a decision of an immigration officer, refusing his application for a work permit. The officer was not satisfied that the applicant would leave Canada at the end of the permit period. The applicant raised issues of procedural fairness and substance. Notably, in its submissions, the applicant argued that Chinook, which is software which apparently permits bulk processing to review applications more quickly, contributed to the decision being procedurally unfair and substantively unreasonable. The applicant said the following in its submissions:
It is not certain as to what information was analysed by the aforesaid software and what was presented to the decision maker to make up a decision. It can be presumed that not enough of human input has gone into it, which is not appropriate for a complicated case involving business immigration. It is also not apt in view of the importance of the decision to the individual, who has committed a great deal of funds for this purpose.
[…]
The test set by the Supreme Court in Vavilov has not been met.
Issue: Is the decision fair and reasonable?
Holding: Yes; see para 24: “However, the evidence is that the Decision was made by a Visa Officer and not by software. I agree the Decision had input assembled by artificial intelligence, but it seems to me the Court on judicial review is to look at the record and the Decision and determine its reasonableness in accordance with Vavilov. Whether a decision is reasonable or unreasonable will determine if it is upheld or set aside, whether or not artificial intelligence was used. To hold otherwise would elevate process over substance.”
Analysis: We are all aware that AI is going to change the world as we know it—one only need look at the sophisticated drawings prepared by AI in mere minutes. Administrative decision-making is no exception, and there is a growing literature on the topic. Canadian immigration officials have developed the Chinook software, and the government maintains that it is not AI software (though in this case the applicant called it AI). I want to write a bit about this issue today: TLDR— AI or algorithmic decision-making of this sort could present significant challenges for decision-making from an administrative justice & law perspective, but applicants will need to access more information about Chinook to make these claims persuasive.
In the case itself, the Court obviously gave short shrift to the arguments about the opaqueness of the software at issue. As Will Tao notes in an excellent and informative piece, this may not have been the case to directly challenge the use of Chinook—it appears to be an open question whether Chinook was used here, even though the Court accepts that AI assisted in the defision. I agree with Mr. Tao that the reasons offered in this case seem more responsive than the sort of boilerplate that some judges of the Federal Court have, rightly, questioned: see Zibadel v Canada (Citizenship and Immigration), 2023 FC 285 (Issue #81). You can judge for yourself: see para 3.
Nonetheless, I think there are both administrative justice and administrative law challenges that one could mount to Chinook and like technology. I offer these tentatively, since there are clearly open questions about what Chinook is, how precisely it is used, and so on. Part of the problem is that while the administrative justice challenges are evident, it is difficult—based on what we know thus far about Chinook—to slot these challenges into cognizable administrative law claims. And given that machine-assisted decision-making is quite unlike other administrative tools considered by the courts—guidelines, consultation procedures, the like—it is an open question whether our administrative law is equipped for what is to come.
Let me start with the administrative justice challenges. Chinook—at the very least—is a bulk processing tool that allows easy access for decision-makers to information relevant to a particular case. But as I have noted in this newsletter before, to the extent that Chinook is related to the phenomenon of boilerplate reasons with which the Federal Court has been dealing, there is a risk that it could undermine the prospect of responsive reasons. Among other things (as the Daponte Affidavit prepared for the Ocran litigation outlines), Chinook “presents standard language that Decision-Makers may select, review, and modify…” The ideal, of course, is that where appropriate, decision-makers will modify the standard language to suit the individual circumstances, as required by the law.
But we should not ignore the context in which Chinook arises. Canada’s immigration system is under severe stress, with backlogs becoming the norm. The government's aim with Chinook is to maximize processing times. This creates the age-old problem: as government tries to maximize “efficiency” in its operations (perhaps a fool’s errand), there could be an inevitable drift away from legal requirements that take some time to meet; ie) writing responsive reasons.
The administrative justice concern is bolstered by what we know about Chinook. As the Daponte Affidavit outlines, Chinook can produce standard template language that can be used by officers, but which they are not compelled to use (see paras 35-36 of the affidavit). But I wonder—how many officers do in fact exercise discretion to modify the standard language? Under conditions of high volume processing, individualizing these template statements might not be a high priority, especially with the sort of cognitive biases that might be introduced. It’d be interesting to know if officers receive training on when it is appropriate to deviate from standard language.
The battle between efficiency and consistency with legal norms is not all or nothing. Chinook, according to the Daponte Affidavit, does lead to better processing times. But there is a risk that what we gain in efficiency we lose in legal consistency.
The challenge becomes putting these intuitions into administrative law claims. One could imagine procedural and substantive challenges. Under the heading of procedural fairness, some of the arguments in Haghshenas echoed concerns that underlie the Consolidated-Bathurst Ltd, [1990] 1 SCR 282 line of cases (see particularly Ellis-Don Ltd v Ontario, 2001 SCC 4). These cases implement the age-old principle of “he who hears must decide.” In other words, as a matter of impartiality, decision-makers cannot farm out the act of decision to someone who has not heard the evidence and arguments. Consolidated-Bathurst, as expressed in Ellis-Don, encapsulates a few basic rules as it pertains to so-called “institutional consultation procedures”—procedures designed to encourage institutional consistency in decision-making while upholding the “he who hears must decide rule.” The idea is, basically, that consultation procedures cannot be imposed from above; and that the decision-maker must retain the ability to find facts without external, coercive pressure.
The Court accepts here that Chinook was used in the case [24]. But the Court is right to accept that the decision itself was made by the officer: this is an application of the well-known distinction between the decision and reasons for decision. Nonetheless, one should ask, somewhat metaphysically: at what point does the officer’s decision become, in reality, the machine’s? This is not a problem that arises with typical soft law instruments, at least not in the sense I’m concerned with here. Much of this turns on how Chinook actually works. If cases with different but related facts are being decided with the same reasons, and the officer is not exercising any discretion of her own, then complex questions are raised from the perspective of institutional impartiality. But as it stands, I’m candidly not sure how Chinook will develop.
There are also administrative law challenges that can arise on substantive grounds ie) (1) fettering and (2) Vavilovian reasonableness. Fettering of discretion is a substantive error that occurs when a decision-maker fastens on to some non-binding guideline rather than the law (see Stemijon Investments Ltd v Canada (Attorney General), 2011 FCA 299 at para 22). Assistance from guidelines of this sort is not forbidden: Vavilov contemplates the use of “training materials, checklists and templates for the purpose of streamlining and strengthening institutional best practices, provided that these methods do not operate to fetter decision making” (Vavilov, at para 130, my emphasis).
The idea that Chinook could fetter discretion is, in theory, not far-fetched, but yet again, we simply do not have all the information. On the face of it, it does appear that officers can modify boilerplate language. Whether they do so or not is an important open question. But the preservation of that ability to modify is important for the purposes of fettering. In strict terms, Chinook standard language is not mandatory.
The reasonableness analysis might be more promising, but it would warrant a broader conception of what it means for a decision to be “reasonable.” As this case shows, judicial review is typically premised around an identifiable decision. Under reasonableness review, of course, that decision is based on a record. Increasingly, under Vavilov, the record has taken on greater importance. Because the record on judicial review is usually the same as the record on before the decision-maker, how records are formed becomes a question that feeds into the overall reasonableness of the decision (if you’re interested, I wrote about this phenomenon here). This is especially so under the Vavilov framework for reasonableness.
In the Federal Courts, efforts to understand how records are formed and who practically made the decision at issue have surrounded Rule 317 of the Federal Courts Rules, which allow an applicant to bring a motion for disclosure for material “relevant to an application that is in possession of a tribunal” and not in possession of the party. But this Rule, admittedly, plays a limited role, and is not a tool of documentary discovery as in a civil action: see Tsleil-Waututh, 2017 FCA 128 at para 115. Yet this provides a potential avenue to understand how it is that Chinook works, and how it impacted a particular decision. This could also be relevant to procedural fairness and the “he who hears must decide” rule: Rule 317 could be broader when there is such an allegation of procedural fairness (see Château d’Ivoire Stores Inc. v. Canada (Attorney General), 2022 FC 405 at para 27).
There are challenges with this. But it appears to me that in order to get the right idea about what Chinook is and does, and how it directly relates to supposed unfairness or unreasonableness in a decision, more information in specific cases could be required. More broadly, it seems to me that the boilerplate that could result from the use of Chinook is a major administrative justice problem, and could be a problem under administrative law. But more needs to be done to demonstrate that this is so.
This puts the already overworked immigration bar in the completely unenviable position of having to understand how this software works. In that sense, this case is a disappointment. That said, given the size of the challenges at issue with Chinook, the Court’s statement that “[w]hether a decision is reasonable will determine if it is upheld or set aside, whether or not artificial intelligence was used” is, at this point, premature [24]. Accordingly, I do not view this case as much of a “precedent” for the legitimacy of Chinook.
See also: Paul Daly.
Labourers International Union of North America, Local 615 v Grafton Developments Inc, 2023 NSCA 25 (April 5, 2023)
Category: Reasonableness review
Context: From the case:
Local 615 of the Labourers International Union applied for certification for construction labourers employed by Grafton Developments Inc. After a hearing, the Labour Board determined that some cleaners were not performing labourers’ construction work and were not in the bargaining unit. Grafton sought judicial review. The judge of the Supreme Court of Nova Scotia held the Labour Board’s Decision offended the reasonableness standard of review. Local 615 appealed to the Court of Appeal.
Issue: Is the decision reasonable? Specifically, while the applicant raised many issues, one of interest was the reviewing judge’s decision that the Board did not grapple with decisions of the Ontario Labour Relations Board raised by the applicant. Is this a flaw under the reasonableness standard?
Holding: The decision is reasonable.
Analysis: It was a valiant attempt by Grafton to cite favourable OLRB decisions that found that workers in comparable circumstances (cleaning) were engaged in “construction work” for the purposes of that province’s legislation [118]. To its credit, the reviewing judge found the argument persuasive. The judge thought the analysis and factors raised in these cases was central to the case, of major importance to the parties, and therefore, the apparent lack of consideration of these cases was fatal.
The Court of Appeal does not accept this argument, for good reason, I think. But I think it actually makes an error that, ultimately, does not affect its result but is important to clarify—it confuses the test for judicial precedents with the one for internal administrative precedents. It cites Vavilov’s para 112, which falls under the “Other Statutory or Common Law” constraint, for the proposition that there is a requirement of bindingness when we are dealing with precedents from other administrative actors. But the Vavilov Court was speaking here about judicial precedents. On administrative precedents, decision-makers are not bound by their previous decisions, but must justify departures from “established” internal precedents (Vavilov, at para 131).
Here, it does not make a difference, as the Court lists other reasons for distinguishing the OLRB cases, particularly that the NS Board applied similar principles anyhow, and that the Ontario legislation is fundamentally different [123]. I agree—there is a risk with importing administrative precedents from other provinces especially dealing with different legislation. But to be nitpicky—in theory—we should keep these (somewhat similar) tests for judicial and administrative precedents distinct. The judicial concept of bindingness is just not applicable in the administrative context when dealing with administrative precedents (see Vavilov, at para 131).
Disclaimer: Nothing in the SEAR is legal advice. The SEAR is not designed to be comprehensive. It is designed to collect some cases that I believe are interesting to practitioners and students, and to provide my quick thoughts. Not all of the cases presented are fully analyzed ie) I’ve selected the most relevant issues in some of the cases. Please complement the SEAR with your own research and study. Any mistakes are my own.