Teresa Scassa - Blog

This is Part III of a series of posts that look at the proposed amendments to Canada’s Artificial Intelligence and Data Act (which itself is still a Bill, currently before the INDU Committee for study). Part I provided a bit of context and a consideration of some of the new definitions in the Bill. Part II looked at the categories of ‘high-impact’ AI that the Bill now proposes to govern. This post looks at the changed role of the AI and Data Commissioner.

The original version of the Artificial Intelligence and Data Act (Part II of Bill C-27) received considerable criticism for its oversight mechanisms. Legal obligations for the ethical and transparent governance of AI, after all, depend upon appropriate oversight and enforcement for their effectiveness. Although AIDA proposed the creation of an AI and Data Commissioner (Commissioner), this was never meant to be an independent regulator. Ultimately, AIDA placed most of the oversight obligations in the hands of the Minister of Industry – the same Minister responsible for supporting the growth of Canada’s AI sector. Critics considered this to be a conflict of interest. A series of proposed amendments to AIDA are meant to address these concerns by reworking the role of the Commissioner.

Section 33(1) of AIDA makes it clear that the AI and Data Commissioner will be a “senior official of the department over which the Minister presides”, and their appointment involves being designated by the Minister. This has not changed, although the amendments would delete from this provision language stating that the Commissioner’s role is “to assist the Minister in the administration and enforcement” of AIDA.

The proposed amendments elevate the Commissioner somewhat, giving them a series of powers and duties, to which the Minister can add through delegation (s. 33(3)). So, for example, it will be the newly empowered Commissioner (Commissioner 2.0) who receives reports from those managing a general-purpose or high impact system where there are reasonable grounds to suspect that the use of the system has caused serious harm (s. 8.2(1)(e), s. 11(1)(g)). Commissioner 2.0 can also order someone managing or making available a general-purpose system to provide them with the accountability framework they are required to create under s. 12 (s. 13(1)) and can provide guidance or recommend corrections to that framework (s. 13(2)). Commissioner 2.0 can compel those making available or managing an AI system to provide the Commissioner with an assessment of whether the system is high impact, and in relation to which subclass of high impact systems set out in the schedule. Commissioner 2.0 can agree or disagree with the assessment, although if they disagree, their authority seems limited to informing the entity in writing with their reasons for disagreement.

More significant are Commissioner 2.0’s audit powers. Under the original version of AIDA, these were to be exercised by the Minister – the powers are now those of the Commissioner (s. 15(1)). Further, Commissioner 2.0 may order (previously this was framed as “require”) that the person either conduct an audit themselves or that the person engage the services of an independent auditor. The proposed amendments also empower the Commissioner to conduct an audit to determine if there is a possible contravention of AIDA. This strengthens the audit powers by ensuring that there is at least an option that is not at least somewhat under the control of the party being audited. The proposed amendments give Commissioner 2.0 additional powers necessary to conduct an audit and to carry out testing of an AI system (s. 15(2.1)). Where Commissioner 2.0 conducts an audit, they must provide the audited party with a copy of the report (s. 15(3.1)) and where the audit is conducted by the person responsible or someone retained by them, they must provide a copy to the Commissioner (s. 15(4)).

The Minister still retains some role with respect to audits. He or she may request that the Commissioner conduct an audit. In an attempt to preserve some independence of Commissioner 2.0, the Commissioner, when receiving such a request, may either carry out the audit or decline to do so on the basis that there are no reasonable grounds for an audit, so long as they provide the Minister with their reasons (s. 15.1(1)(b)). The Minister may also order a person to take actions to bring themselves into compliance with the law (s. 16) or to cease making available or terminate the operation of a system if the Minister considers compliance to be impossible (s. 16(b)) or has reasonable grounds to believe that the use of the system “gives rise to a risk of imminent and serious harm” (s. 17(1)).

As noted above, Commissioner 2.0 (a mere employee in the Minister’s department) will have order making powers under the amendments. This is something the Privacy Commissioner of Canada, an independent agent of Parliament, appointed by the Governor in Council, is hoping to get in Bill C-27. If so, it will be for the first time since the enactment of PIPEDA in 2000. Orders of Commissioner 2.0 or the Minister can become enforceable as orders of the Federal Court under s. 20.

Commissioner 2.0 is also empowered to share information with a list of federal or provincial government regulators where they have “reasonable grounds to believe that the information may be relevant to the administration or enforcement by the recipient of another Act of Parliament or of a provincial legislature.” (s. 26(1)). Reciprocally, under a new provision, federal regulators may also share information with the Commissioner (s. 26.1). Additionally, Commissioner 2.0 may “enter into arrangements” with different federal regulators and/or the Ministers of Health and Transport in order to assist those actors with the “exercise of their powers or the performance of their functions and duties” in relation to AI (s. 33.1). These new provisions strengthen a more horizontal, multi-regulator approach to governing AI which is an improvement in the Bill, although this might eventually need to be supplemented by corresponding legislative amendments – and additional funding – to better enable the other commissioners to address AI-related issues that fit within their areas of competence.

The amendments also impose upon Commissioner 2.0 a new duty to report on the administration and enforcement of AIDA – such a report is to be “published on a publicly available website”. (s. 35.1) The annual reporting requirement is important as it will increase transparency regarding the oversight and enforcement of AIDA. For his or her part, the Minister is empowered to publish information, where it is in the public interest, regarding any contravention of AIDA or where the use of a system gives rise to a serious risk of imminent harm (ss. 27 and 28).

Interestingly, AIDA, which provides for the potential imposition of administrative monetary penalties for contraventions of the Act does not indicate who is responsible for setting and imposing these penalties. Section 29(1)(g) makes it clear that “the persons or classes of persons who may exercise any power, or perform any duty or function, in relation to the [AMP] scheme” is left to be articulated in regulations.

The AIDA also makes it an offence under s. 30 for anyone to obstruct or provide false or misleading information to “the Minister, anyone acting on behalf of the Minister or an independent auditor in the exercise of their powers or performance of their duties or functions under this Part.” This remains unchanged from the original version of AIDA. Presumably, since Commissioner 2.0 would exercise a great many of the oversight functions, this is meant to apply to the obstruction or misleading of the Commissioner – but it will only do so if the Commissioner is characterized as someone “acting on behalf of the Minister”. This is not language of independence, but then there are other features of AIDA that also counter any view that even Commissioner 2.0 is truly independent (and I mean others besides the fact that they are an employee under the authority of the Minister and handpicked by the Minister). Most notable of these is that should the Commissioner become incapacitated or absent, or should they simply never be designated by the Minister, it is the Minister who will exercise their powers and duties (s. 33(4)).

In sum, then, the proposed amendments to AIDA attempt to give some separation between the Minister and Commissioner 2.0 in terms of oversight and enforcement. At the end of the day, however, Commissioner 2.0 is still the Minister’s hand-picked subordinate. Commissioner 2.0 does not serve for a specified term and has no security of tenure. In their absence, the Minister exercises their powers. It falls far short of independence.

My previous post looked at some of the new definitions in the proposed amendments to the Artificial Intelligence and Data Act (AIDA) which is Part III of Bill C-27. These include a definition of “high impact” AI, and a schedule of classes of high-impact AI (the Schedule is reproduced at the end of this post). The addition of the schedule changes AIDA considerably, and that is the focus of this post.

The first two classes in the Schedule capture contexts that can clearly affect individuals. Class 1 addresses AI used in most aspects of employment, and Class 2 relates to the provision of services. On the provision of services (which could include things like banking and insurance), the wording signals that it will apply to decision-making about the provision of services, their cost, or the prioritization of recipients. To be clear, AIDA does not prohibit systems with these functions. They are simply characterized as “high impact” so that they will be subject to governance obligations. A system to determine creditworthiness can still reject individuals; and companies can still prioritize preferred customers – as long as the systems are sufficiently transparent, free from bias and do not cause harm.

There is, however, one area which seems to fall through the cracks of Classes 1 & 2: rental accommodation. A lease is an interest in land – it is not a service. Human rights legislation in Canada typically refers to accommodation separately from services for this reason. AI applications are already being used to screen and select tenants for rental accommodation. In the midst of a housing crisis, this is surely an area that is high-impact and where the risks of harm from flawed AI to individuals and families searching for a place to live are significant. This gap needs to be addressed – perhaps simply by adding “or accommodation” after each use of the term “service” in Class 2.

Class 3 rightly identifies biometric systems as high risk. It also includes systems that use biometrics in “the assessment of an individual’s behaviour or state of mind.” Key to the scope of this section will be the definition of “biometric”. Some consider biometric data to be exclusively physiological data (fingerprints, iris scans, measurements of facial features, etc.). Yet others include behavioral data in this class if it is used for the second identified purpose – the assessment of behaviour or state of mind. Behavioural data, though, is potentially a very broad category. It can include data about a person’s gait, or their speech or keystroke patterns. Cast even more broadly, it could include things such as “geo-location and IP addresses”, “purchasing habits”, “patterns of device use” or even “browser history and cookies”. If that is the intention behind Class 3, then conventional biometric AI should be Part One of this class; Part Two should be the use of an AI system to assess an individual’s behaviour or state of mind (without referring specifically to biometrics in order to avoid confusion). This would also, importantly, capture the highly controversial area of AI for affect recognition. It would be unfortunate if the framing of the class as ‘biometrics’ led to an unduly narrow interpretation of the kind of systems or data involved. The explanatory note in the Minister’s cover letter for this provision seems to suggest (although it is not clear) that it is purely physiological biometric data that is intended for inclusion and not a broader category. If this is so, then Class 3 seems unduly narrow.

Class 4 is likely to be controversial. It addresses content moderation and the prioritization and presentation of content online and identifies these as high-impact algorithmic activities. Such systems are in widespread use in the online context. The explanatory note from the Minister observes that such systems “have important potential impacts on Canadians’ ability to express themselves, as well as pervasive effects at societal scale” (at p. 4). This is certainly true although the impact is less direct and obvious than the impact of a hiring algorithm, for example. Further, although an algorithm that presents a viewer of online streaming services with suggestions for content could have the effect of channeling a viewer’s attention in certain directions, it is hard to see this as “high impact” in many contexts, especially since there are multiple sources of suggestions for online viewing (including word of mouth). That does not mean that feedback loops and filter bubbles (especially in social media) do not contribute to significant social harms – but it does make this high impact class feel large and unwieldy. The Minister’s cover letter indicates that each of the high-impact classes presents “distinct risk profiles and consequently will require distinct risk management strategies.” (at p. 2). Further, he notes that the obligations that will be imposed “are intended to scale in proportion to the risks they present. A low risk use within a class would require correspondingly minimal mitigation effort.” (at p. 2). Much will clearly depend on regulations.

Class 5 relates to the use of AI in health care or emergency services, although it explicitly excludes medical devices because these are already addressed by Health Canada (which recently consulted on the regulation of AI-enabled medical devices). This category also demonstrates some of the complexity of regulating AI in Canada’s federal system. Many hospital-based AI technologies are being developed by researchers affiliated with the hospitals and who are not engaged in the interprovincial or international trade and commerce which is necessary for AIDA to apply. AIDA will only apply to those systems developed externally and in the context of international or interprovincial trade and commerce. While this will still capture many applications, it will not capture all – creating different levels of governance within the same health care context.

It is also not clear what is meant, in Class 5, by “use of AI in matters relating to health care”. This could be interpreted to mean health care that is provided within what is understood as the health care system. Understood more broadly, it could extend to health-related apps – for example, one of the many available AI-enabled sleep trackers, or an AI-enabled weight loss tool (to give just two examples). I suspect that what is intended is the former, even though, with health care in crisis and more people turning to alternate means to address their health issues, health-related AI technologies might well deserve to be categorized as high-impact.

Class 6 involves the use of an AI system by a court or administrative body “in making a determination in respect of an individual who is a party to proceedings before the court or administrative body.” In the first place, this is clearly not meant to apply to automated decision-making generally – it seems to be limited to judicial or quasi-judicial contexts. Class 6 must also be reconciled with s. 3 of AIDA, which provides that AIDA does not apply “with respect to a government institution as defined in s. 3 of the Privacy Act.” This includes the Immigration and Refugee Board, for example, as well as the Canadian Human Rights Commission, the Parole Board, and the Veterans Review and Appeal Board. Making sense of this, then, it would be the tools used by courts or tribunals and developed or deployed in the course of interprovincial or international trade and commerce that would be considered high impact. The example given in the Minister’s letter seems to support this – it is of an AI system that provides an assessment of “risk of recidivism based on historical data” (at p. 5).

However, Class 6 is confusing because it identifies the context rather than the tools as high impact. Note that the previous classes address the use of AI “in matters relating to” the subject matter of the class, whereas class 6 identifies actors – the use of AI by a court or tribunal. There is a different focus. Yet the same tools used by courts and tribunals might also be used by administrative bodies or agencies that do not hold hearings or that are otherwise excluded from the application of AIDA. For example, in Ewert v. Canada, the Supreme Court of Canada considered an appeal by a Métis man who challenged the use of recidivism-risk assessment tools by Correctional Services of Canada (to which AIDA would not apply according to s. 3). If this type of tool is high-risk, it is so whether it is used by Correctional Services or a court. This suggests that the framing of Class 6 needs some work. It should perhaps be reworded to identify tools or systems as high impact if they are used to determine the rights, entitlements or status of individuals.

Class 7 addresses the use of an AI system to assist a peace officer “in the exercise and performance of their law enforcement powers, duties and function”. Although “peace officer” receives the very broad interpretation found in the Criminal Code, that definition is modified in the AIDA by language that refers to the exercise of specific law enforcement powers. This should still capture the use of a broad range of AI-enabled tools and technologies. It is an interesting question whether AIDA might apply more fulsomely to this class of AI systems (not just those developed in the course of interprovincial or international trade) as it might be considered to be rooted in the federal criminal law power.

AIDA Schedule:

High-Impact Systems — Uses

1. The use of an artificial intelligence system in matters relating to determinations in respect of employment, including recruitment, referral, hiring, remuneration, promotion, training, apprenticeship, transfer or termination.

2. The use of an artificial intelligence system in matters relating to

(a) the determination of whether to provide services to an individual;

(b) the determination of the type or cost of services to be provided to an individual; or

(c) the prioritization of the services to be provided to individuals.

3. The use of an artificial intelligence system to process biometric information in matters relating to

(a) the identification of an individual, other than in cases in which the biometric information is processed with the individual’s consent to authenticate their identity; or

(b) the assessment of an individual’s behaviour or state of mind.

4. The use of an artificial intelligence system in matters relating to

(a) the moderation of content that is found on an online communications platform, including a search engine or social media service; or

(b) the prioritization of the presentation of such content.

5. The use of an artificial intelligence system in matters relating to health care or emergency services, excluding a use referred to in any of paragraphs (a) to (e) of the definition device in section 2 of the Food and Drugs Act that is in relation to humans.

6. The use of an artificial intelligence system by a court or administrative body in making a determination in respect of an individual who is a party to proceedings before the court or administrative body.

7. The use of an artificial intelligence system to assist a peace officer, as defined in section 2 of the Criminal Code, in the exercise and performance of their law enforcement powers, duties and functions.

Note: This is the first in a series of posts that will look at the proposed amendments to Canada's Artificial Intelligence and Data Act, which is Part III of Bill C-27, currently before Parliament. The amendments are extensive and have only just been introduced. Please consider these assessments to be preliminary.

Canada’s Artificial Intelligence and Data Act (AIDA) (Part III of Bill C-27) has passed second reading and is currently before the INDU Committee for study. Early in this committee process, the Minister of Industry Philippe Champagne announced that his department was working on amendments to AIDA in response to considerable criticism. Those amendments have now been tabled for consideration by the committee.

One of the criticisms of the Bill was that it left almost all of its substance to be developed in regulations. It is unsurprising, then, that the amendments are almost as long as the original bill. While it is certainly the case that the amendments contain more detail than the original text, some of the additional length is attributable to new provisions intended to address generative AI systems. This highlights just how quickly things are moving in the AI space, as generative AI was not on anyone’s legislative radar when Bill C-27 was introduced in June 2022.

One of the criticisms of AIDA was the absence of any specific prior consultation before its appearance in Bill C-27. This, combined with its lack of substance on many issues, raised basic concerns about how it would apply and to what. For example, AIDA was to govern “high-impact” AI systems, but the definition of such systems was left to regulations. Concerns were also raised about oversight being largely in the hands of the Minister of Industry who is also responsible for supporting Canada’s AI sector.

The proposed amendments demonstrate that ISED has been listening to the feedback it has received since June 2022, just as it has been adapting to the challenges of generative AI, and engaging with its international partners on AI governance issues. The amendments, which include new definitions, more explicit obligations, and governance principles for generative AI, will make AIDA a better bill. They may be enough to garner sufficient support to pass it into law, something which the Minister describes as “pivotal”.

This is the first in a series of posts that will explore some of the changes proposed to AIDA – as well as some of the remaining issues. This post addresses some of the new definitions.

The proposed AIDA amendments propose a new definition of “artificial intelligence system” which would read: “a technological system that, using a model, makes inferences in order to generate output, including predictions, recommendations or decisions” (s. 2). This provides greater alignment with the OECD definition of an AI system (“An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment.”) It is an improvement over the previous definition, which was criticized for being too specific about the types of techniques used in AI. It is unclear, though, why the new AIDA definition does not include “content” as an output as is the case with the OECD definition. The AIDA definition is also supplemented by a separate definition for a “general-purpose system”, which is “an artificial intelligence system that is designed to be adapted for use, in many fields and for many purposes and activities, including fields, purposes and activities not contemplated during the system’s development” (s. 5(1)). There is a further definition for a “machine learning model”, which is “a digital representation of patterns identified in data through the automated processing of the data using an algorithm designed to enable the recognition or replication of those patterns”. A new s. 5(2) makes it clear that the definition of AI system includes general-purpose systems, and that general-purpose systems can also be high-impact. These new definitions reflect the major changes in both the technology and in the evolving regulatory context in the short time since AIDA was introduced. They also shape a new framework for obligations under the legislation.

The proposed amendments also contain a definition of “high-impact system”: “an artificial intelligence system of which at least one of the intended uses may reasonably be concluded to fall within a class of uses set out in the schedule”. (s. 5(1)). The previous version of AIDA left the articulation of “high impact” to future regulations. The schedule sets out a list of classes that describe certain uses. These are:

High-Impact Systems — Uses

1. The use of an artificial intelligence system in matters relating to determinations in respect of employment, including recruitment, referral, hiring, remuneration, promotion, training, apprenticeship, transfer or termination.

2. The use of an artificial intelligence system in matters relating to

(a) the determination of whether to provide services to an individual;

(b) the determination of the type or cost of services to be provided to an individual; or

(c) the prioritization of the services to be provided to individuals.

3. The use of an artificial intelligence system to process biometric information in matters relating to

(a) the identification of an individual, other than in cases in which the biometric information is processed with the individual’s consent to authenticate their identity; or

(b) the assessment of an individual’s behaviour or state of mind.

4. The use of an artificial intelligence system in matters relating to

(a) the moderation of content that is found on an online communications platform, including a search engine or social media service; or

(b) the prioritization of the presentation of such content.

5. The use of an artificial intelligence system in matters relating to health care or emergency services, excluding a use referred to in any of paragraphs (a) to (e) of the definition device in section 2 of the Food and Drugs Act that is in relation to humans.

6. The use of an artificial intelligence system by a court or administrative body in making a determination in respect of an individual who is a party to proceedings before the court or administrative body.

7. The use of an artificial intelligence system to assist a peace officer, as defined in section 2 of the Criminal Code, in the exercise and performance of their law enforcement powers, duties and functions.

(Note: the classes in this schedule will be the subject of the next blog post)

The list is not intended to be either closed or permanent. Under a proposed s. 36.1, the Governor in Council (GinC) can enact regulations amending the schedule by adding, modifying, or deleting a category of use. Any such decision by the GinC is to be guided by criteria set out in s. 36.1. These include the risks of adverse impact on “the economy or any other aspect of Canadian society and on individuals, including on individual’s health and safety and on their rights recognized in international human rights treaties to which Canada is a party”. The GinC must also consider the “severity and extent” of any adverse impacts, as well as the “social and economic circumstances of any individuals who may experience” such impacts. A final consideration is whether the uses in the category are adequately addressed under another Act of Parliament or of a provincial legislature.

The AIDA only applies to “high impact” systems, and since there is no screening or registration process, it is up to those who manage or make such systems available to identify them as such and to meet the obligations set out in the law. A proposed s. 14 would empower the AI and Data Commissioner to order a person who makes available or who manages an AI system to provide the Commissioner with their assessment of whether the system is a high impact system, a general purpose system (which can also be high impact), or a machine learning model intended to be incorporated into a high impact system.

My next post will look at the classes of “high-impact” AI as set out in the Schedule.

The following is a short excerpt from a new paper which looks at the public sector use of private sector personal data (Teresa Scassa, “Public Sector Use of Private Sector Personal Data: Towards Best Practices”, forthcoming in (2024) 47:2 Dalhousie Law Journal ) The full pre-print version of the paper is available here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4538632

Governments seeking to make data-driven decisions require the data to do so. Although they may already hold large stores of administrative data, their ability to collect new or different data is limited both by law and by practicality. In our networked, Internet of Things society, the private sector has become a source of abundant data about almost anything – but particularly about people and their activities. Private sector companies collect a wide variety of personal data, often in high volumes, rich in detail, and continuously over time. Location and mobility data, for example, are collected by many different actors, from cellular service providers to app developers. Financial sector organizations amass rich data about the spending and borrowing habits of consumers. Even genetic data is collected by private sector companies. The range of available data is constantly broadening as more and more is harvested, and as companies seek secondary markets for the data they collect.

Public sector use of private sector data is fraught with important legal and public policy considerations. Chief among these is privacy since access to such data raises concerns about undue government intrusion into private lives and habits. Data protection issues implicate both public and private sector actors in this context, and include notice and consent, as well as data security. And, where private sector data is used to shape government policies and actions, important questions about ethics, data quality, the potential for discrimination, and broader human rights questions also arise. Alongside these issues are interwoven concerns about transparency, as well as necessity and proportionality when it comes to the conscription by the public sector of data collected by private companies.

This paper explores issues raised by public sector access to and use of personal data held by the private sector. It considers how such data sharing is legally enabled and within what parameters. Given that laws governing data sharing may not always keep pace with data needs and public concerns, this paper also takes a normative approach which examines whether and in what circumstances such data sharing should take place. To provide a factual context for discussion of the issues, the analysis in this paper is framed around two recent examples from Canada that involved actual or attempted access by government agencies to private sector personal data for public purposes. The cases chosen are different in nature and scope. The first is the attempted acquisition and use by Canada’s national statistics organization, Statistics Canada (StatCan), of data held by credit monitoring companies and financial institutions to generate economic statistics. The second is the use, during the COVID-19 pandemic, of mobility data by the Public Health Agency of Canada (PHAC) to assess the effectiveness of public health policies in reducing the transmission of COVID-19 during lockdowns. The StatCan example involves the compelled sharing of personal data by private sector actors; while the PHAC example involves a government agency that contracted for the use of anonymized data and analytics supplied by private sector companies. Each of these instances generated significant public outcry. This negative publicity no doubt exceeded what either agency anticipated. Both believed that they had a legal basis to gather and/or use the data or analytics, and both believed that their actions served the public good. Yet the outcry is indicative of underlying concerns that had not properly been addressed.

Using these two quite different cases as illustrations, the paper examines the issues raised by the use of private sector data by government. Recognizing that such practices are likely to multiply, it also makes recommendations for best practices. Although the examples considered are Canadian and are shaped by the Canadian legal context, most of the issues they raise are of broader relevance. Part I of this paper sets out the two case studies that are used to tease out and illustrate the issues raised by public sector use of private sector data. Part II discusses the different issues and makes recommendations.

The full pre-print version of the paper is available here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4538632

The government of the United Kingdom has published a consultation paper seeking input into its proposal for AI regulation. The paper is aptly titled A pro-innovation approach to AI regulation, since it restates that point insistently throughout the document. The UK proposal provides an interesting contrast to Canada’s AI governance bill currently before Parliament.

Both Canada and the UK set out to regulate AI systems with the twin goals of supporting innovation on the one hand, and building trust in AI on the other. (Note here that the second goal is to build trust in AI, not to protect the public. Although the protection of the public is acknowledged as one way to build trust, there is a subtle distinction here). However, beyond these shared goals, the proposals are quite different. Canada’s approach in Part 3 of Bill C-27 (the Artificial Intelligence and Data Act (AIDA)) is to create a framework to regulate as yet undefined “high impact” AI. The definition of “high impact” as well as many other essential elements of the bill are left to be articulated in regulations. According to a recently published companion document to the AIDA, leaving so much of the detail to regulations is how the government proposes to keep the law ‘agile’ – i.e. capable of responding to a rapidly evolving technological context. The proposal would also provide some governance for anonymized data by imposing general requirements to document the use of anonymized personal information in AI innovation. The Minister of Innovation is made generally responsible for oversight and enforcement. For example, the AIDA gives the Minister of Innovation the authority (eventually) to impose stiff administrative monetary penalties on bad actors. The Canadian approach is similar to that in the EU AI Act in that it aims for a broad regulation of AI technologies, and it chooses legislation as the vehicle to do so. It is different in that the EU AI Act is far more detailed and prescriptive; the AIDA leaves the bulk of its actual legal requirements to be developed in regulations.

The UK proposal is notably different from either of these approaches. Rather than create a new piece of legislation and/or a new regulatory authority, the UK proposes to set out five principles for responsible AI development and use. Existing regulators will be encouraged and, if necessary, specifically empowered, to regulate AI according to these principles within their spheres of regulatory authority. Examples of regulators who will be engaged in this framework include the Information Commissioner’s Office, regulators for human rights, consumer protection, health care products and medical devices, and competition law. The UK scheme also accepts that there may need to be an entity within government that can perform some centralized support functions. These may include monitoring and evaluation, education and awareness, international interoperability, horizon scanning and gap analysis, and supporting testbeds and sandboxes. Because of the risk that some AI technologies or issues may fall through the cracks between existing regulatory schemes, the government anticipates that regulators will assist government in identifying gaps and proposing appropriate actions. These could include adapting the mandates of existing regulators or providing new legislative measures if necessary.

Although Canada’s federal government has labelled its approach to AI regulation as ‘agile’, it is clear that the UK approach is much closer to the concept of agile regulation. Encouraging existing regulators to adapt the stated AI principles to their remit and to provide guidance on how they will actualize these principles will allow them to move quickly, so long as there are no obvious gaps in legal authority. By contrast, even once passed, it will take at least two years for Canada’s AIDA to have its normative blanks filled in by regulations. And, even if regulations might be somewhat easier to update than statutes, guidance is even more responsive, giving regulators greater room to manoeuvre in a changing technological landscape. Embracing the precepts of agile regulation, the UK scheme emphasizes the need to gather data about the successes and failures of regulation itself in order to adapt as required. On the other hand, while empowering (and resourcing) existing regulators will have clear benefits in terms of agility, the regulatory gaps could well be important ones – with the governance of large language models such as ChatGPT as one example. While privacy regulators are beginning to flex their regulatory muscles in the direction of ChatGPT, data protection law will only address a subset of the issues raised by this rapidly evolving technology. In Canada, AIDA’s governance requirements will be specific to risk-based regulation of AI, and will apply to all those who design, develop or make AI systems available for use (unless of course they are explicitly excluded under one of the many actual and potential exceptions).

Of course, the scheme in the AIDA may end up as more of a hybrid between the EU and the UK approaches in that the definition of “high impact” AI (to which the AIDA will apply) may be shaped not just by the degree of impact of the AI system at issue but also by the existence of other suitable regulatory frameworks. In other words, the companion document suggests that some existing regulators (health, consumer protection, human rights, financial institutions) have already taken steps to extend their remit to address the use of AI technologies within their spheres of competence. In this regard, the companion document speaks of “regulatory gaps that must be filled” by a statute such as AIDA as well as the need for the AIDA to integrate “seamlessly with existing Canadian legal frameworks”. Although it is still unclear whether the AIDA will serve only to fill regulatory gaps, or will provide two distinct layers of regulation in some cases, one of the criteria for identifying what constitutes a “high impact” system includes “[t]he degree to which the risks are adequately regulated under another law”. The lack of clarity in the Canadian approach is one of its flaws.

There is a certain attractiveness in the idea of a regulatory approach like that proposed by the UK – one that begins with existing regulators being both specifically directed and further enabled to address AI regulation within their areas of responsibility. As noted earlier, it seems far more agile than Canada’s rather clunky bill. Yet such an approach is much easier to adopt in a unitary state than in a federal system such as Canada’s. In Canada, some of the regulatory gaps are with respect to matters otherwise under provincial jurisdiction. Thus, it is not so simple in Canada to propose to empower and resource all implicated regulators, nor is it as easy to fill gaps once they are identified. These regulators and the gaps between them might fall under the jurisdiction of any one of 13 different governments. The UK acknowledges (and defers) its own challenges in this regard with respect to devolution at paragraph 113 of its white paper, where it states: “We will continue to consider any devolution impacts of AI regulation as the policy develops and in advance of any legislative action”. Instead, the AIDA, Canada leverages its general trade and commerce power in an attempt to provide AI governance that is as comprehensive as possible. It isn’t pretty (since it will not capture all AI innovation that might have impacts on people) but it is part of the reality of the federal state (or the state of federalism) in which we find ourselves.

The federal government’s proposed Artificial Intelligence and Data Act (AIDA) is currently before Parliament as part of Bill C-27, a bill that will also reform Canada’s private sector data protection law. The AIDA, which I have discussed in more detail in a series of blog posts (here, here, and here), has been criticized for being a shell of a law with essential components (including the definition of the “high impact AI” to which it will apply) being left to as-yet undrafted regulations. The paucity of detail in the AIDA, combined with the lack of public consultation, has prompted considerable frustration and concern from AI developers and from civil society alike. In response to these concerns, the government published, on March 13, 2023, a companion document that explains the government’s thinking behind the AIDA. The document is a useful read as it makes clear some of the rationales for different choices that have been made in the bill. It also obliquely engages with many of the critiques that have been leveled at the AIDA. Unlike a consultation document, however, where feedback is invited to improve what is being proposed, the companion document is essentially an apology (in the Greek sense of the word) – something that is written in defense or explanation. At this stage, any changes will have to come as amendments to the bill.

Calling this a ‘companion document’ also somewhat tests the notion of “companion”, since it was published nine months after the AIDA was introduced in Parliament in June 2022. The document explains that the government seeks to take “the first step towards a new regulatory system designed to guide AI innovation in a positive direction, and to encourage the responsible adoption of AI technologies by Canadians and Canadian businesses.” The AIDA comes on the heels of the European Union’s draft AI Act – a document that is both more comprehensive and far more widely consulted upon. Pressure on Canada to regulate AI is heightened by the activity in the EU. This is evident in the introduction to the companion document, which speaks of the need to work with international partners to achieve global protection for Canadians and to ensure that “Canadian firms can be recognized internationally as meeting robust standards.”

An important critique of the AIDA has been that it will apply only to “high impact” AI. By contrast, the EU AI Act sets a sliding scale of obligations, with the most stringent obligations applying to high risk applications, and minimal obligations for low risk AI. In the AIDA companion document, there is no explanation of why the AIDA is limited to high impact AI. The government explains that defining the scope of the Act in regulations will allow for greater precision, as well as for updates as technology progresses. The companion document offers some clues about what the government considers relevant to determining whether an AI system is high-impact. Factors include the type of harm, the severity of harm, and the scale of use. Although this may help understand the concept of high impact, it does not explain why governance was only considered for high and not medium or low impact AI. This is something that cannot be fixed by the drafting of regulations. The bill would have to be specifically amended to provide for governance for AI with different levels of impact according to a sliding scale of obligations.

Another important critique of the AIDA has been that it unduly focuses on individual rather than collective or broader harms. As the US’s NIST AI Risk Management Framework aptly notes, AI technologies “pose risks that can negatively impact individuals, groups, organizations, communities, society, the environment and the planet” (at p. 1). The AIDA companion document addresses this critique by noting that the bill is concerned both with individual harms and with systemic bias (defined as discrimination). Yet, while it is crucially important to address the potential for systemic bias in AI, this is not the only collective harm that should be considered. The potential for AI to be used to generate and spread disinformation or misinformation, for example, can create a different kind of collective harm. Flawed AI could potentially also result in environmental damage that is the concern of all. The companion document does little to address a broader notion of harm – but how can it? The AIDA specifically refers to and defines “individual harm”, and also addresses biased output as discriminatory within the meaning of the Canadian Human Rights Act. Only amendments to the bill can broaden its scope to encompass other forms of collective harm. Such amendments are essential.

Another critique of the AIDA is that it relies for its oversight on the same Ministry that is responsible for promoting and supporting AI innovation in Canada. The companion document tackles this concern, citing the uniqueness of the AI context, and stating that “administration and enforcement decisions have important implications for policy”, such that oversight and the encouragement of innovation “would need to be [sic] work in close collaboration in the early years of the framework under the direction of the Minister.” The Minister will be assisted by a Ministry staffer who will be designated the AI and Data Commissioner. The document notes that the focus in the early days of the legislation will be on helping organizations become compliant: “The Government intends to allow ample time for the ecosystem to adjust to the new framework before enforcement actions are undertaken.” The ample time will include the (at least) two years before the necessary regulations are drafted (though note that if some key regulations are not drafted, the law will never take effect), as well as any subsequent ‘adjustment’ time. Beyond this, the document is quite explicit that compliance and enforcement should not get unnecessarily in the way of the industry. The AIDA contains other mechanisms, including requiring companies to hire their own auditors for audits and having an appointed Ministerial advisory committee to reassure those who remain concerned about governance. Yet these measures do nothing to address a core lack of independent oversight. This lack is particularly noteworthy given that the same government has proposed the creation of an ill-advised Personal Information and Data Protection Tribunal (in Part II of Bill C-27) in order to establish another layer between the Privacy Commissioner and the enforcement of Bill C-27’s proposed Consumer Privacy Protection Act. It is difficult to reconcile the almost paranoid approach taken to the Privacy Commissioner’s role with the in-house, “we’re all friends here” approach to AI governance in the AIDA. It is hard to see how this lack of a genuine oversight framework can be fixed without a substantial rewrite of the bill.

And that brings us to the reality that we must confront with this bill: AI technologies are rapidly advancing and are already having significant impacts on our lives. The AIDA is deeply flawed, and the lack of consultation is profoundly disturbing. Yet, given the scarcity of space on Parliament’s agenda and the generally fickle nature of politics, the failure of the AIDA could lead to an abandonment of attempts to regulate in this space – or could very substantially delay them. As debate unfolds over the AIDA, Parliamentarians will have to ask themselves the unfortunate question of whether the AIDA is unsalvageable, or whether it can be sufficiently amended to be better than no law at all.

Artificial intelligence (AI) is already being used to assist government decision-making, although we have little case law that explores issues of procedural fairness when it comes to automated decision systems. This is why a recent decision of the Federal Court is interesting. In Barre v. Canada (Citizenship and Immigration) two women sought judicial review of a decision of the Refugee Protection Division (RPD) which had stripped them of their refugee status. They raised procedural fairness issues regarding the possible reliance upon an AI tool – in this case facial recognition technology (FRT). The case allows us to consider some procedural fairness guideposts that may be useful where evidence derived from AI-enabled tools is advanced.

The Decision of the Refugee Protection Division

The applicants, Ms Barre and Ms Hosh, had been granted refugee status after advancing claims related to their fear of sectarian and gender-based violence in their native Somalia. The Minister of Public Safety and Emergency Preparedness (the Minister) later applied under s. 109 of the Immigration and Refugee Protection Act to have that decision vacated on the basis that it was “obtained as a result of directly or indirectly misrepresenting or withholding material facts relating to a relevant matter”.

The Minister had provided the RPD with photos that compared Ms Barre and Ms Hosh the applicants) with two Kenyan women who had been admitted to Canada on student visas shortly before Ms Barre and Ms Hosh filed their refugee claims (the claims were accepted in 2017). The applicants argued that the photo comparisons relied upon by the Minister had been made using Clearview AI’s facial recognition service built upon scraped images from social media and other public websites. The Minister objected to arguments and evidence about Clearview AI, maintaining that there was no proof that this service had been used. Clearview AI had ceased providing services in Canada on 6 July 2020, and the RPD accepted the Minister’s argument that it had not been used, finding that “[a]n App that is banned to operate in Canada would certainly not be used by a law enforcement agency such as the CBSA” (at para 7). The Minister had also argued that it did not have to disclose how it arrived at the photo comparisons because of s. 22 of the Privacy Act, and the RPD accepted this assertion.

The photo comparisons were given significant weight in the RPD’s decision to overturn the applicants’ refugee status. The RPD found that there were “great similarities” between the photos of the Kenyan students and the applicants, and concluded that they were the same persons. The RPD also considered notes in the Global Case Management System to the effect that the Kenyan students did not attend classes at the school where they were enrolled. In addition, the CBSA submitted affidavits indicating that there was no evidence that the applicants had entered Canada under their own names. The RPD concluded that the applicants were Kenyan citizens who had misrepresented their identity in the refugee proceedings. It found that these factual misrepresentations called into question the credibility of their allegations of persecution. It also found that, since they were Kenyan, they had not advanced claims against their country of nationality in the refugee proceedings, as required by law. The applicants sought judicial review of the decision to revoke their refugee status, arguing that it was unreasonable and breached their rights to procedural fairness.

Judicial Review

Justice Go of the Federal Court ruled that the decision was unreasonable for a number of reasons. A first error was allowing the introduction of the photo comparisons into evidence “without requiring the Minister to disclose the methodology used in procuring the evidence” (at para 31). The Minister had invoked s. 22 of the Privacy Act, but Justice Go noted that there were many flaws with the Minister’s reliance on s. 22. Section 22 is an exception to an individual’s right of access to their personal information. Justice Go noted that the applicants were not seeking access to their personal information; rather, they were making a procedural fairness argument about the photo comparisons relied upon by the Minister and sought information about how the comparisons had been made. Section 22(2), which was specifically relied upon by the Minister, allows a request for disclosure of personal information to be refused on the basis that it was “obtained or prepared by the Royal Canadian Mounted Police while performing policing services for a province or municipality…”, and this circumstance simply was not relevant.

Section 22(1)(b), which was not specifically argued by the Minister, allows for a refusal to disclose personal information where to do so “could reasonably be expected to be injurious to the enforcement of any law of Canada or a province or the conduct of lawful investigations…” Justice Go noted that case law establishes that a court will not support such a refusal on the basis that because there is an investigation, harm from disclosure can be presumed. Instead, the head of an institution must demonstrate a “nexus between the requested disclosure and a reasonable expectation of probable harm” (at para 35, citing Canadian Association of Elizabeth Fry Societies v. Canada). Exceptions to access rights must be given a narrow interpretation, and the burden of demonstrating that a refusal to disclose is justifiable lies with the head of the government institution. Justice Go also noted that “the Privacy Act does not operate “so as to limit access to information to which an individual might be entitled as a result of other legal rules or principles”” (at para 42) such as, in this case, the principles of procedural fairness.

Justice Go found that the RPD erred by not clarifying what ‘personal information’ the Minister sought to protect; and by not assessing the basis for the Minister’s 22 arguments. She also noted that the RPD had accepted the Minister’s bald assertions that the CBSA did not rely on Clearview AI. Even if the company had ceased offering its services in Canada by July 6, 2020, there was no evidence regarding the date on which the photo comparisons had been made. Justice Go noted that the RPD failed to consider submissions by the applicants regarding findings by the privacy commissioners of Canada, BC, Alberta and Quebec regarding Clearview AI and its activities, as well as on the “danger of relying on facial recognition software” (at para 46).

The Minister argued that even if its s. 22 arguments were misguided, it could still rely upon evidentiary privileges to protect the details of its investigation. Justice Go noted that this was irrelevant in assessing the reasonableness of the RPD’s decision, since such arguments had not been made before or considered by the RPD. She also observed that when parties seek to exempt information from disclosure in a hearing, they are often required at least to provide it to the decision-maker to assess. In this case the RPD did not ask for or assess information on how the investigation had been conducted before deciding that information about it should not be disclosed. She noted that: “The RPD’s swift acceptance of the Minister’s exemption request, in the absence of a cogent explanation for why the information is protected from disclosure, appears to be a departure from its general practice” (at para 55).

Justice Go also observed that information about how the photo comparisons were made could well have been relevant to the issues to be determined by the RPD. If the comparisons were generated through use of FRT – whether it was using Clearview AI or the services of another company – “it may call into question the reliability of the Kenyan students’ photos as representing the Applicants, two women of colour who are more likely to be misidentified by facial recognition software than their white cohorts as noted by the studies submitted by the Applicants” (at para 56). No matter how the comparisons were made – whether by a person or by FRT technology – some evidence should have been provided to explain the technique. Justice Go found it unreasonable for the RPD to conclude that the evidence was reliable simply based upon the Minister’s assertions.

Justice Go also found that the RPD’s conclusion that the applicants were, in fact, the two Kenyan women, was unreasonable. Among other things, she found that the decision “failed to provide adequate reasons for the RPD’s conclusion that the two Applicants and the two Kenyan students were the same persons based on the photo comparisons” (at para 69). She noted that although the RPD referenced ‘great similarities’ between the women in the two sets of photographs, there were also some marked dissimilarities which were not addressed. There simply was no adequate explanation as to how the conclusion was reached that the applicants were the Kenyan students.

The decision of the RPD was quashed and remitted to be reconsidered by a differently constituted panel of the RPD.

Ultimately, Justice Go sends a clear message that the Minister cannot simply advance photo comparison evidence without providing an explanation for how that evidence was derived. At the very least, then, there is an obligation to indicate whether an AI technology was used in the decision-making process. Even if there is some legal basis for shielding the details of the Minister’s methods of investigation, there may still need to be some disclosure to the decision-maker regarding the methods used. Justice Go’s decision is also a rebuke of the RPD which accepted the Minister’s evidence on faith and asked no questions about its methodology or probity. In her decision, Justice Go take serious note of concerns about accuracy and bias in the use of FRT, particularly with racialized individuals, and it is clear that these concerns heighten the need for transparency. The decision is important for setting some basic standards to meet when it comes to reviewing evidence that may have been derived using AI. It is also a sobering reminder that those checks and balances failed at first instance – and in a high stakes context.

This is the third in my series of posts on the Artificial Intelligence and Data Act (AIDA) found in Bill C-27, which is part of a longer series on Bill C-27 generally. Earlier posts on the AIDA have considered its purpose and application, and regulated activities. This post looks at the harms that the AIDA is designed to address.

The proposed Artificial Intelligence and Data Act (AIDA), which is the third part of Bill C-27, sets out to regulate ‘high-impact’ AI systems. The concept of ‘harm’ is clearly important to this framework. Section 4(b) of the AIDA states that a purpose of the legislation is “to prohibit certain conduct in relation to artificial intelligence systems that may result in serious harm to individuals or harm to their interests”.

Under the AIDA, persons responsible for high-impact AI systems have an obligation to identify, assess, and mitigate risks of harm or biased output (s. 8). Those persons must also notify the Minister “as soon as feasible” if a system for which they are responsible “results or is likely to result in material harm”. There are also a number of oversight and enforcement functions that are triggered by harm or a risk of harm. For example, if the Minister has reasonable grounds to believe that a system may result in harm or biased output, he can demand the production of certain records (s. 14). If there is a serious risk of imminent harm, the Minister may order a person responsible to cease using a high impact system (s. 17). The Minister is also empowered to make public certain information about a system where he believes that there is a serious risk of imminent harm and the publication of the information is essential to preventing it (s. 28). Elevated levels of harm are also a trigger for the offence in s. 39, which involves “knowing or being reckless as to whether the use of an artificial intelligence system is likely to cause serious physical or psychological harm to an individual or substantial damage to an individual’s property”.

‘Harm’ is defined in s. 5(1) to mean:

(a) physical or psychological harm to an individual;

(b) damage to an individual’s property; or

(c) economic loss to an individual.

I have emphasized the term “individual” in this definition because it places an important limit on the scope of the AIDA. First, it is unlikely that the term ‘individual’ includes a corporation. Typically, the word ‘person’ is considered to include corporations, and the word ‘person’ is used in this sense in the AIDA. This suggests that “individual” is meant to have a different meaning. The federal Interpretation Act is silent on the issue. It is a fair interpretation of the definition of ‘harm’ that “individual” is not the same as “person”, and means an individual (human) person. The French version uses the term “individu”, and not “personne”. The harms contemplated by this legislation are therefore to individuals and not to corporations.

Defining harm in terms of individuals has other ramifications. The AIDA defines high-risk AI systems in terms of their impacts on individuals. Importantly, this excludes groups and communities. It also very significantly focuses on what are typically considered quantifiable harms, and uses language that suggests quantifiability (economic loss, damage to property, physical or psychological harm). Some important harms may be difficult to establish or to quantify. For example, class action lawsuits relating to significant data breaches have begun to wash up on the beach of lost causes due to the impossibility of proving material loss either because, although thousands may have been impacted, the individual losses are impossible to quantify, or because it is impossible to prove a causal link between very real identity theft and that particular data breach. Consider an AI system that manipulates public opinion through an algorithm that drives content to individuals based on its shock value rather than its truth. Say this happens during a pandemic and it convinces people that they should not get vaccinated or take other recommended public health measures. Say some people die because they were misled in this way. Say other people die because they were exposed to infected people who were misled in this way. How does one prove the causal link between the physical harm of injury or death of an individual and the algorithm? What if there is an algorithm that manipulates voter sentiment in a way that changes the outcome of an election? What is the quantifiable economic loss or psychological harm to any individual? How could causation be demonstrated? The harm, once again, is collective.

The EU AI Act has also been criticized for focusing on individual harm, but the wording of that law is still broader than that in the AIDA. The EU AI Act refers to high-risk systems in terms of “harm to the health and safety or a risk of adverse impact on fundamental rights of persons”. This at least introduces a more collective dimension, and it avoids the emphasis on quantifiability.

The federal government’s own Directive on Automated Decision-Making (DADM) which is meant to guide the development of AI used in public sector automated decision systems (ADS) also takes a broader approach to impact. In assessing the potential impact of an ADS, the DADM takes into account: “the rights of individuals or communities”, “the health or well-being of individuals or communities”, “the economic interests of individuals, entities, or communities”, and “the ongoing sustainability of an ecosystem”.

With its excessive focus on individuals, the AIDA is simply tone deaf to the growing global understanding of collective harm caused by the use of human-derived data in AI systems.

One response of the government might be to point out that the AIDA is also meant to apply to “biased output”. Biased output is defined in the AIDA as:

content that is generated, or a decision, recommendation or prediction that is made, by an artificial intelligence system and that adversely differentiates, directly or indirectly and without justification, in relation to an individual on one or more of the prohibited grounds of discrimination set out in section 3 of the Canadian Human Rights Act, or on a combination of such prohibited grounds. It does not include content, or a decision, recommendation or prediction, the purpose and effect of which are to prevent disadvantages that are likely to be suffered by, or to eliminate or reduce disadvantages that are suffered by, any group of individuals when those disadvantages would be based on or related to the prohibited grounds. (s. 5(1)) [my emphasis]

The argument here will be that the AIDA will also capture discriminatory biases in AI. However, I have underlined the part of this definition that once again returns the focus to individuals, rather than groups. It can be very hard for an individual to demonstrate that a particular decision discriminated against them (especially if the algorithm is obscure). In any event, biased AI will tend to replicate systemic discrimination. Although it will affect individuals, it is the collective impact that is most significant – and this should be recognized in the law. The somewhat obsessive focus on individual harm in the AIDA may unwittingly help perpetuate denials of systemic discrimination.

It is also important to note that the definition of “harm” does not include “biased output”, and while the terms are used in conjunction in some cases (for example, in s. 8’s requirement to “identify, assess and mitigate the risks of harm or biased output”), other obligations relate only to “harm”. Since the two are used conjunctively in some parts of the statute, but not others, a judge interpreting the statute might presume that when only one of the terms is used, then it is only that term that is intended. Section 17 of the AIDA allows the Minister to order a person responsible for a high-impact system to cease using it or making it available if there is a “serious risk of imminent harm”. Section 28 permits the Minister to order the publication of information related to an AI system where there are reasonable grounds to believe that the use of the system gives rise to “a serious risk of imminent harm”. In both cases, the defined term ‘harm’ is used, but not ‘biased output’.

The goals of the AIDA to protect against harmful AI are both necessary and important, but in articulating the harm that it is meant to address, the Bill underperforms.

As part of my series on Bill C-27, I will be writing about both the proposed amendments to Canada’s private sector data protection law and the part of the Bill that will create a new Artificial Intelligence and Data Act (AIDA). So far, I have been writing about privacy, and my posts on consent, de-identification, data-for-good, and the right of erasure are already available. Posts on AIDA, will follow, although I still have a bit more territory on privacy to cover first. However, in the meantime, as a teaser, perhaps you might be interested in playing a bit of statutory MadLibs…...

Have you ever played MadLibs? It’s a paper-and-pencil game where someone asks the people in the room to supply a verb, noun, adverb, adjective, or body part, and the provided words are used to fill in the blanks in a story. The results are often absurd and sometimes hilarious.

The federal government’s proposal in Bill C-27 for an Artificial Intelligence and Data Act, really lends itself to a game of statutory MadLibs. This is because some of the most important parts of the bill are effectively left blank – either the Minister or the Governor-in-Council is tasked in the Bill with filling out the details in regulations. Do you want to play? Grab a pencil, and here goes:

Company X is developing an AI system that will (insert definition of ‘high impact system). It knows that this system is high impact because (insert how a company should assess impact). Company X has established measures to mitigate potential harms by (insert measures the company took to comply with the regulations) and has also recorded (insert records it kept), and published (insert information to be published).

Company X also had its system audited by an auditor who is (insert qualifications). Company X is being careful, because if it doesn’t comply with (insert a section of the Act for which non-compliance will count as a violation), it could be found to have committed a (insert degree of severity) violation. This could lead to (insert type of proceeding).

Company X, though, will be able to rely on (insert possible defence). However, if (insert possible defence) is unsuccessful, Company X may be liable to pay an Administrative Monetary Penalty if they are a (insert category of ‘person’) and if they have (insert factors to take into account). Ultimately, if they are unhappy with the outcome, they can launch a (insert a type of appeal proceeding).

Because of this regulatory scheme, Canadians can feel (insert emotion) at how their rights and interests are protected.

Note: The following is my response to the call for submissions on the recommendations following the third review of Canada’s Directive on Automated Decision-Making. Comments are due by June 30, 2022. If you are interested in commenting, please consult the Review Report and the Summary of Key Issues and Proposed Amendments. Comments can be sent to This e-mail address is being protected from spambots. You need JavaScript enabled to view it .

The federal Directive on Automated Decision-Making (DADM) and its accompanying Algorithmic Impact Assessment tool (AIA) are designed to provide governance for the adoption and deployment of automated decision systems (ADS) by Canada’s federal government. Governments are increasingly looking to ADS in order to speed up routine decision-making processes and to achieve greater consistency in decision-making. At the same time, there are reasons to be cautious. Automated decision systems carry risks of incorporating and replicating discriminatory bias. They may also lack the transparency required of government decision-making, particularly where important rights or interests are at stake. The DADM, which has been in effect since April 2019 (with compliance mandatory no later than April 2020), sets out a series of obligations related to the design and deployment of automated decision-making systems. The extent of the obligations depends upon a risk assessment, and the AIA is the tool by which the level of risk of the system is assessed.

Given that this is a rapidly evolving area, the DADM provides that it will be reviewed every six months. It is now in its third review. The first two reviews led to the clarification of certain obligations in the DADM and to the development of guidelines to aid in its interpretation. This third review proposes a number of more substantive changes. This note comments on some of these changes and proposes an issue for future consideration.

Clarify and Broaden the Scope

A key recommendation in this third round of review relates to the scope of the DADM. Currently, the DADM applies only to ‘external’ services of government – in other words services offered to individuals or organizations by government. It does not apply internally. This is a significant gap when one considers the expanding use of ADS in the employment context. AI-enabled decision systems have been used in hiring processes, and they can be used to conduct performance reviews, and to make or assist in decision-making about promotions and internal workforce mobility. The use of AI tools in the employment context can have significant impacts on the lives and careers of employees. It seems a glaring oversight to not include such systems in the governance regime for ADM. The review team has recommended expanding the scope of the DADM to include internal as well as external services. They note that this move would also extend the DADM to any ADS used for “grants and contributions, awards and recognition, and security screening” (Report at 11). This is an important recommendation and one which should be implemented.

The review team also recommends a clarification of the language regarding the application of the DADM. Currently it puts within its scope “any system, tool, or statistical models used to recommend or make an administrative decision about a client”. Noting that “recommend” could be construed as including only those systems that recommend a specific outcome, as opposed to systems that process information on behalf of a decision-maker, the team proposes replacing “recommend” with “support”. This too is an important recommendation which should be implemented.

Periodic Reviews

Currently the DADM provides for its review every six months. This was always an ambitious review schedule. No doubt it was motivated by the fact that the DADM was a novel tool designed to address a rapidly emerging and evolving technology with potentially significant implications. The idea was to ensure that it was working properly and to promptly address any issues or problems. In this third review, however, the team recommends changing the review period from six months to two years. The rationale is that the six-month timetable makes it challenging for the team overseeing the DADM (which is constantly in a review cycle), and makes it difficult to properly engage stakeholders. They also cite the need for the DADM to “display a degree of stability and reliability, enabling federal institutions and the clients they serve to plan and act with a reasonable degree of confidence.” (Report at 12).

This too is a reasonable recommendation. While more frequent reviews were important in the early days of the DADM and the AIA, reviews every six months seem unduly burdensome once initial hiccups are resolved. A six-month review cycle engages the team responsible for the DADM in a constant cycle of review, which may not be the best use of resources. The proposed two-year review cycle would allow for a more experience to be garnered with the DADM and AIA, enabling a more substantive assessment of issues arising. Further, a two-year window is much more realistic if stakeholders are to be engaged in a meaningful way. Being asked to comment on reports and proposed changes every six months seems burdensome for anyone – including an already stretched civil society sector. The review document suggests that Canada’s Chief Information Officer could request completion of an off-cycle review if the need arose, leaving room for the possibility that a more urgent issue could be addressed outside of the two-year review cycle.

Data Model and Governance

The third review also proposes amendments to provide for what it describes as a more ‘holistic’ approach to data governance. Currently, the DADM focuses on data inputs – in other words on assessing the quality, relevance and timeliness of the data used in the model. The review report recommends the addition of an obligation to establish “measures to ensure that data used and generated by the Automated Decision System are traceable, protected, and appropriately retained and disposed of in accordance with the Directive on Service and Digital, Directive on Privacy Practices, and Directive on Security Management”. It will also recommend amendments to extend testing and assessment beyond data to underlying models, in order to assess both data and algorithms for bias or other problems. These are positive amendments which should be implemented.

Explanation

The review report notes that while the DADM requires “meaningful explanations” of how automated decisions were reached, and while guidelines provide some detail as to what is meant by explainability, there is still uncertainty about what explainability entails. The Report recommends adding language in Appendix C, in relation to impact assessment, that will set out the information necessary for ‘explainability’. This includes:

• The role of the system in the decision-making process;
  
• The training and client data, their source and method of collection, if applicable;
• The criteria used to evaluate client data and the operations applied to process it; and
• The output produced by the system and any relevant information needed to interpret it in the context of the administrative decision.

Again, this recommendation should be implemented.

Reasons for Automation

The review would also require those developing ADM systems for government to specifically identify why it was considered necessary or appropriate to automate the existing decision-making process. The Report refers to a “clear and demonstrable need”. This is an important additional criterion as it requires transparency as to the reasons for automation – and that these reasons go beyond the fact that vendor-demonstrated technologies look really cool. As the authors of the review note, requiring justification also helps to assess the parameters of the system adopted – particularly if the necessity and proportionality approach favoured by the Office of the Privacy Commissioner of Canada is adopted.

Transparency

The report addresses several issues that are relevant to the transparency dimensions of the DADM and the accompanying AIA. Transparency is an important element of the DADM, and it is key both to the legitimacy of the adoption of ADS by government, but also to its ongoing use. Without transparency in government decision-making that impacts individuals, organizations and communities, there can be no legitimacy. There are a number of transparency elements that are built into the DADM. For example, there are requirements to provide notice of automated decision systems, a right to an explanation of decisions that is tailored to the impact of the decision, and a requirement not just to conduct an AIA, but to publish the results. The review report includes a number of recommendations to improve transparency. These include a recommendation to clarify when an AIA must be completed and released, greater transparency around peer review results, more explicit criteria for explainability, and adding additional questions to the AIA. These are all welcome recommendations.

At least one of these recommendations may go some way to allaying my concerns with the system as it currently stands. The documents accompanying the report (slide 3 of summary document) indicate that there are over 300 AI projects across 80% of federal institutions. However, at the time of writing, only four AIAs were published on the open government portal. There is clearly a substantial lag between development of these systems and release of the AIAs. The recommendation that an AIA be not just completed but also released prior to the production of the system is therefore of great importance to ensuring transparency.

It may be that some of the discrepancy in the numbers is attributable to the fact that the DADM came into effect in 2020, and it was not grandfathered in for projects already underway. For transparency’s sake, I would also recommend that a public register of ADS be created that contains basic information about all government ADS. This could include their existence and function, as well as some transparency regarding explainability, the reasons for adoption, and measures taken to review, assess and ensure the reliability of these systems. Although it is too late, in the case of these systems, to perform a proactive AIA, there should be some form of reporting tool that can be used to provide important information, for transparency purposes, to the public.

Consideration for the Future

The next review of the DADM and the AIA should also involve a qualitative assessment of the AIAs that have been published to date. If the AIA is to be a primary tool not just for assessing ADS but for providing transparency about them, then they need to be good. Currently there is a requirement to conduct an AIA for a system within the scope of the DADM – but there is no explicit requirement for it to be of a certain quality. A quick review of the four AIAs currently available online shows some discrepancy between them in terms of the quality of the assessment. For example, the project description for one such system is an unhelpful 9-word sentence that does not make clear how AI is actually part of the project. This is in contrast to another that describes the project in a 14-line paragraph. These are clearly highly divergent in terms of the level of clarity and detail provided.

The first of these two AIAs also seems to contain contradictory answers to the AIA questionnaire. For example, the answer to the question “Will the system only be used to assist a decision-maker” is ‘yes’. Yet the answer to the question “Will the system be replacing a decision that would otherwise be made by a human” is also ‘yes’. Either one of these answers is incorrect, or the answers do not capture how the respondent interpreted these questions. These are just a few examples. It is easy to see how use of the AIA tool can range from engaged to pro forma.

The obligations imposed on departments with respect to ADS vary depending upon the risk assessment score. This score is evaluated through the questionnaire, and one of the questions asks “Are clients in this line of business particularly vulnerable?” In the AIA for an access to information (ATIP) tool, the answer given to this question is “no”. Of course, the description of the tool is so brief that it is hard to get a sense of how it functions. However, I would think that the clientele for an ATIP portal would be quite diverse. Some users will be relatively sophisticated (e.g., journalists or corporate users). Others will be inexperienced. For some of these, information sought may be highly important to them as they may be seeking access to government information to right a perceived wrong, to find out more about a situation that adversely impacts them, and so on. In my view, this assessment of the vulnerability of the clients is not necessarily accurate. Yet the answer provided contributes to a lower overall score and thus a lower level of accountability. My recommendation for the next round of reviews is to assess the overall effectiveness of the AIA tool in terms of the information and answers provided and in terms of their overall accuracy.

I note that the review report recommends adding questions to the AIA in order to improve the tool. Quite a number of these are free text answers, which require responses to be drafted by the party completing the AIA. Proposed questions include ones relating to the user needs to be addressed, how the system will meet those needs, and the effectiveness of the system in meeting those needs, along with reasons for this assessment. Proposed questions will also ask whether non-AI-enabled solutions were also considered, and if so, why AI was chosen as the preferred method. A further question asks what the consequences would be of not deploying the system. This additional information is important both to assessing the tool and to providing transparency. However, as noted above, the answers will need to be clear and sufficiently detailed in order to be of any use.

The AIA is crucial to assessing the level of obligation and to ensuring transparency. If AIAs are pro forma or excessively laconic, then the DADM can be as finely tuned as can be, but it will still not achieve desired results. The review committee’s recommendation that plain language summaries of peer review assessments also be published will provide a means of assessing the quality of the AIAs, and thus it is an important recommendation to strengthen both transparency and compliance.

A final issue that I would like to address is that, to achieve transparency, people will need to be able to easily find and access the information about the systems. Currently, AIAs are published on the Open Government website. There, they are listed alphabetically by title. This is not a huge problem right now, since there are only four of them. As more are published, it would be helpful to have a means of organizing them by department or agency, or by other criteria (including risk/impact score) to improve their findability and usability. Further, it will be important that any peer review summaries are linked to the appropriate AIAs. In addition to publication on the open government portal, links to these documents should be made available from department, agency or program websites. It would also be important to have an index or registry of AI in the federal sector – including not just those projects covered by the DADM, but also those in production prior to the DADM’s coming into force.

[Note: I have written about the DADM and the AIA from an administrative law perspective. My paper, which looks at the extent to which the DADM addresses administrative law concerns regarding procedural fairness, can be found here.]