The AI Register’s Hidden Assumptions

Author: Denis Avetisyan

Canada’s attempt to publicly document its AI systems reveals more about bureaucratic priorities than algorithmic risks.

Government of Canada organizations demonstrate a varied adoption of artificial intelligence systems, with deployment concentrated across specific provinces and territories-a distribution that has evolved over time alongside shifting system statuses-suggesting a dynamic and geographically uneven integration of the technology within the public sector.

An analysis of the Canadian AI Register demonstrates how its structure constructs a limited view of algorithmic accountability by obscuring crucial details of bureaucratic processes and discretionary practices.

While governments increasingly champion algorithmic transparency through initiatives like public AI registers, these tools often present a constrained view of complex bureaucratic realities. This paper, ‘Bureaucratic Silences: What the Canadian AI Register Reveals, Omits, and Obscures’, analyzes Canada’s Federal AI Register, revealing how it constructs a specific ontology of AI focused on ‘reliable tooling’ rather than ‘contestable decision-making’ through selective documentation of 409 systems. Our analysis, utilizing the ADMAPS framework, demonstrates a systematic obscuring of human discretion, training procedures, and uncertainty management inherent in algorithmic governance. Does this emphasis on technical descriptions over sociotechnical context ultimately automate accountability into a performative exercise, offering visibility without genuine contestability?

Mapping the Algorithmic State: Peering into the Canadian AI Register

Canada’s national artificial intelligence strategy is built on a dual commitment: fostering rapid technological advancement and ensuring responsible deployment. This approach recognizes that the transformative potential of AI is inextricably linked to public trust, demanding a proactive focus on transparency and accountability. While innovation is actively encouraged through funding and research initiatives, the strategy explicitly emphasizes the need to mitigate potential risks associated with algorithmic bias, data privacy, and the erosion of due process. This commitment isn’t merely aspirational; it’s reflected in efforts to document AI systems within government, aiming to illuminate how these technologies are shaping public services and impacting citizens’ lives. The underlying principle is that a thriving AI ecosystem requires not only ingenuity but also a robust framework for ethical oversight and public understanding.

The Canadian AI Register represents a foundational effort to bring visibility to the increasing use of artificial intelligence within the federal government. Currently detailing 409 AI systems deployed across 42 departments and agencies, the register encompasses a broad spectrum of applications, from emerging Generative AI tools to established Predictive Systems utilized in areas like resource allocation and service delivery. This comprehensive cataloging serves not only as an inventory of current AI adoption, but also as a critical step towards fostering greater transparency and accountability in governmental decision-making processes increasingly influenced by algorithmic technologies. By documenting these deployments, the register lays the groundwork for informed public discourse and responsible innovation in the realm of artificial intelligence within the Canadian public sector.

Despite the Canadian AI Register’s scope – cataloging 409 AI systems across federal departments – a critical gap in transparency emerges upon closer examination. The register reveals what researchers term ‘Bureaucratic Silences,’ indicating a lack of detailed information surrounding how these algorithms actually function and impact citizens. While the register represents a significant first step, only 4% of the 303 automated tools currently listed include publicly available Algorithmic Impact Assessments. This scarcity of assessments hinders meaningful public scrutiny and accountability, suggesting that the register, in its current form, provides a broad overview of AI deployment without fully illuminating the intricacies of algorithmic decision-making within the Canadian government.

A word cloud visually represents the range of technical capabilities demonstrated by the AI systems, with word size indicating prevalence.

Unpacking the Black Box: Dependencies and the Illusion of Autonomy

Analysis of the Canadian AI Register indicates a substantial level of infrastructural dependence within deployed algorithmic systems. Specifically, 38.1% of registered systems rely on external vendors for core functionality or maintenance. This reliance extends to transnational entities, raising concerns regarding national AI sovereignty and potential vulnerabilities related to supply chain security and data governance. The prevalence of external vendor dependence suggests limited domestic capacity in key areas of AI development and deployment, necessitating further investigation into the nature of these dependencies and their implications for Canadian autonomy in the field.

Despite perceptions of automation, human discretion consistently influences the application of algorithmic systems. Analysis reveals that algorithms rarely function with complete autonomy; instead, human actors make crucial decisions regarding data input, system configuration, interpretation of outputs, and exception handling. This human involvement, however, is frequently under-documented in system specifications and operational procedures, creating a gap between the technical presentation of algorithmic objectivity and the reality of human-in-the-loop operation. The lack of transparency regarding these discretionary practices poses challenges for accountability, auditability, and the effective governance of algorithmic systems.

The ADMAPS (Algorithms, Data, Methods, Actors, Politics, Systems) Framework provides a structured approach to analyzing algorithmic systems within bureaucratic contexts by deconstructing them into constituent parts and examining their interrelationships. Application of the framework consistently reveals significant documentation gaps, or ‘silences’, regarding the rationale for design choices, the sources and quality of data used for training, and the roles of human actors in interpreting and overriding algorithmic outputs. These silences hinder comprehensive impact assessments and accountability mechanisms, limiting understanding of how algorithms shape policy outcomes and potentially exacerbate existing inequalities within bureaucratic processes. Addressing these documentation deficiencies is crucial for ensuring transparency, fairness, and responsible deployment of algorithmic systems in governance.

Towards Sovereign AI: Building Trust Through Transparency and Accountability

Effective AI governance relies significantly on public trust, which is directly correlated with the level of transparency surrounding algorithmic systems. This transparency necessitates clear communication regarding how AI systems are designed, the data they utilize, and the potential societal impacts of their deployment. Without demonstrable openness regarding these factors, public acceptance and confidence in AI technologies will likely erode, hindering responsible innovation and widespread adoption. Establishing mechanisms for auditing algorithms, explaining decision-making processes, and addressing potential biases are crucial components in fostering this necessary public trust and ensuring accountable AI governance.

The Pan-Canadian AI Strategy, while heavily invested in research and development to advance artificial intelligence capabilities, requires concurrent development of robust accountability mechanisms. Currently, the strategy focuses on fostering innovation and economic growth through AI, but without clearly defined methods for addressing potential harms or ensuring responsible deployment, public trust will be eroded. This necessitates the implementation of auditing procedures, impact assessments, and clear lines of responsibility for AI system outcomes. Prioritizing accountability alongside innovation is crucial for establishing ethical and legally sound AI governance frameworks within Canada, and for maintaining public confidence in the technology’s benefits.

The Canadian Sovereign AI Compute Strategy addresses the need to diminish reliance on external infrastructure for artificial intelligence processing and establish domestic control over critical AI resources. Current implementation, as evidenced by the AI Systems Register, demonstrates a pronounced internal focus; 86.3% of developed AI systems are designed for use by Government of Canada (GC) employees. This data indicates that initial deployments prioritize internal government operations and efficiency gains, rather than widespread public-facing applications or commercial use cases, and underscores the strategy’s current emphasis on securing infrastructure to support internal governmental functions.

Beyond Documentation: Reframing Uncertainty and Shaping the Algorithmic Future

The prevailing impulse to eliminate uncertainty from artificial intelligence systems is ultimately counterproductive; instead, responsible AI deployment necessitates its explicit acknowledgement. AI, by its very nature, operates within complex, unpredictable environments and relies on probabilistic models, meaning absolute certainty is rarely, if ever, achievable. Viewing uncertainty not as a flaw, but as an inherent characteristic, allows for the development of adaptive governance frameworks capable of responding to unforeseen consequences and evolving system behaviors. This shift in perspective fosters proactive risk management, encourages ongoing monitoring, and promotes transparency regarding the limitations of AI, building public trust and ensuring accountability – crucial elements for the sustainable and ethical integration of these powerful technologies into society.

Ontological design, a deliberate shaping of AI systems’ foundational assumptions about what exists and what matters, fundamentally influences which aspects of these technologies are governed and for whom accountability is established. Rather than treating AI as a neutral tool simply reflecting pre-existing realities, this approach recognizes that AI actively constructs categories, values, and relationships through its data processing and decision-making logic. Consequently, a proactive ontological design isn’t merely about identifying potential biases, but about strategically defining the very terms of recognition and responsibility within the system itself. By carefully considering which entities, attributes, and connections are encoded – and which are excluded – developers and policymakers can shape not only how AI operates, but also what it deems relevant, ultimately impacting which harms are prevented and whose interests are served.

Research into the Canadian AI Register reveals that its function extends beyond simple documentation; it actively defines what constitutes accountability within public sector artificial intelligence systems. The Register doesn’t merely record responsible AI practices, but instead, through the details it demands and omits – the ‘bureaucratic silences’ – it shapes perceptions of what is considered answerable and what is not. Detailed documentation and consistent monitoring are therefore critical, not just for transparency, but for proactively establishing a framework where AI systems are genuinely held responsible. This process highlights that seemingly neutral bureaucratic tools possess the power to construct, reinforce, or even obscure lines of accountability, making ongoing scrutiny and comprehensive data capture essential for fostering public trust and ensuring ethical AI governance.

The analysis dissects the Canadian AI Register, revealing how seemingly neutral bureaucratic tools actively construct algorithmic accountability rather than simply reflecting it. This echoes Andrey Kolmogorov’s assertion: “The most important thing in science is not to be afraid of new ideas.” The Register, in its attempt to categorize and define AI systems, necessarily imposes a particular ontology, highlighting certain features while silencing others – a process akin to intellectual rule-breaking. The paper demonstrates how this ‘ontological design’ isn’t merely a technical challenge, but a fundamentally political one, shaping how algorithmic governance is understood and practiced. It’s a deliberate act of definition, a construction of reality through categorization, and the study exposes the inherent limitations of attempting to fully capture the complexities of bureaucratic processes within such a framework.

What’s Next?

The Canadian AI Register, as a formalized attempt at algorithmic accountability, presents a curious case. It functions, predictably, as a boundary object – defining not just what is considered accountable, but more importantly, what remains conveniently outside the frame. The analysis suggests the next challenge isn’t simply filling the register with more data, but interrogating the very logic of its categories. What happens if one deliberately mis-categorizes an AI system, forcing a confrontation with the register’s inherent limitations? Or, more provocatively, what if the bureaucratic silences – the omissions regarding discretionary power and operational uncertainty – are not bugs, but core features of any governance scheme?

Future research should abandon the pursuit of ‘complete’ transparency. Such a goal is not only naive but actively obscures the crucial role of interpretation and judgment. Instead, attention should turn to mapping the distribution of those silences. Where do they cluster? Which actors benefit from them? The ADMAPS framework, while useful, offers only a starting point. A more radical approach would treat the register itself as an active agent, a system that shapes the reality it claims to reflect.

Ultimately, the exercise isn’t about better algorithmic governance. It’s about understanding the inherent paradox of trying to codify control. Any register, no matter how meticulously designed, will inevitably create new blind spots. The real work lies in systematically inducing those failures, in deliberately breaking the rules to reveal the underlying mechanisms of power and discretion.

Original article: https://arxiv.org/pdf/2604.15514.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

Mapping the Algorithmic State: Peering into the Canadian AI Register

Unpacking the Black Box: Dependencies and the Illusion of Autonomy

Towards Sovereign AI: Building Trust Through Transparency and Accountability

Beyond Documentation: Reframing Uncertainty and Shaping the Algorithmic Future

What’s Next?

See also: