High-Risk AI Systems and the Problem of Identity in the European AI Act

Technical Summary: High-Risk AI Systems and the Problem of Identity in the European AI Act

Problem Statement
The European Union Artificial Intelligence Act (AIA) establishes a lifecycle governance regime for high-risk AI systems centered on ex-ante conformity assessment, post-market monitoring, and re-assessment upon "substantial modification." This regulatory framework implicitly relies on judgments of "AI identity": regulators and providers must determine when an updated system remains the "same" system over time (diachronic identity) and when contemporaneous instantiations (e.g., copies or variants) should be treated as the same system for regulatory comparability (synchronic identity).

The paper identifies two critical gaps in the AIA's handling of these identity questions:

1. Diachronic Ambiguity: While the AIA defines "substantial modification" as a trigger for re-certification, it lacks a precise, auditable criterion for determining when a system has ceased to be the same entity, particularly regarding the quantification of "appropriate functioning" over time.
2. Synchronic Gap: The AIA provides no internal, auditable criterion for synchronic identity for stand-alone high-risk AI systems. It largely defers these determinations to sectoral harmonization instruments (e.g., medical device regulations), which often rely on vague proxies like "equivalence" or "essential similarity." This creates regulatory blind spots in procurement, liability, and market surveillance where authorities must decide if compliance evidence for one instantiation generalizes to another.

Methodology
The paper adopts a conceptual and auditing approach rather than an empirical or implementation study. It utilizes the function+ framework of artifact identity (Carrara and Vermaas, 2009; Ferrario, 2025) as a metaphysical lens to analyze the AIA's regulatory logic.

The methodology involves:

• Metaphysical Alignment: Mapping the AIA's lifecycle obligations (conformity assessment, monitoring, substantial modification) to the function+ components: techno-function (intended purpose), trustworthiness profile (context-sensitive criteria of appropriate functioning), and trustworthiness level (quantified satisfaction of the profile over time).
• Case Analysis: Applying this framework to two illustrative cases: (1) a clinical triage system undergoing retraining and fairness constraint updates (diachronic challenge), and (2) two biometric border control systems with identical core functionality but different performance optimizations (synchronic challenge).
• Gap Analysis: Comparing the AIA's explicit provisions against the requirements of the function+ framework to identify where the regulation lacks operational precision.
• Workflow Construction: Developing a minimal decision flow (audit workflow) to operationalize synchronic identity judgments.

Key Contributions
The paper makes three primary contributions:

1. *Conceptual Alignment of AIA and function+:* The authors demonstrate that the AIA's lifecycle governance implicitly operationalizes the diachronic component of the function+ framework.

   • Ex-ante conformity assessment fixes the identity-relevant baseline (intended purpose and trustworthiness profile).
   • Post-market monitoring tracks the trustworthiness level over time.
   • Substantial modification acts as the legal discontinuity condition.
     The paper clarifies that while the AIA provides a sufficient condition for re-certification (the emergence of substantial modification), the function+ framework provides necessary and sufficient conditions for persistence, explicitly incorporating the quantification of trustworthiness levels which the AIA currently leaves underspecified.

2. Identification and Resolution of the Synchronic Identity Gap: The paper argues that the AIA fails to provide an internal criterion for synchronic identity for stand-alone systems. It proposes the SI_function+ criterion as a solution: two systems are identical at time t if they share the same techno-function, trustworthiness profile, and trustworthiness level.

   • This criterion anchors comparability in intended function and trustworthiness requirements, making identity decisions inspectable in governance settings like procurement and liability.
   • A minimal audit workflow is proposed (Figure 2), requiring justification of sameness at the conformity assessment stage (purpose/profile) and maintenance of sameness during post-market monitoring (level at time t).

3. Implementation-Facing Recommendations: To render the function+ approach operationally legible within the AIA, the authors propose:

   • Precise Reporting of Intended Purpose: Moving beyond broad descriptions to document the AI functional taxonomy (techno-function, use-function, social function), particularly for generative AI, including explicit exclusions and non-intended uses.
   • Standardized, Auditable Trustworthiness Reporting: Replacing vague "equivalence" claims with standardized reporting of trustworthiness contracts and trustworthiness level functions. The authors suggest using multi-dimensional piecewise-constant functions (e.g., mapping accuracy and safety to "plateaus" of trustworthiness) to aggregate diverse metrics into a composite indicator. This allows for tolerating minor fluctuations in single dimensions while triggering re-assessment only when a system crosses a threshold into a lower trustworthiness plateau.

Results and Claims
The paper does not present experimental results or validated compliance procedures. Instead, its "results" are conceptual clarifications and structural proposals:

• Claim 1: The AIA's lifecycle mechanism implicitly operationalizes diachronic identity, but the function+ framework clarifies this by explicitly linking intended purpose to a quantifiable trustworthiness level, making the "substantial modification" threshold more inspectable.
• Claim 2: The AIA leaves a significant gap in synchronic identity for stand-alone systems. The function+ framework fills this gap by providing a testable criterion (SI_function+) that links intended function and trustworthiness profiles to regulatory comparability.
• Operational Output: The paper provides a "correspondence map" (Figure 1) aligning AIA obligations with metaphysical identity questions and a "minimal audit checklist" (Table 1) for synchronic identity decisions, specifying exactly which fields in AIA artifacts (technical documentation, risk management files, monitoring plans) must be audited.

Significance
The paper positions itself as a conceptual and auditing lens rather than a prescriptive legal reform. Its significance lies in:

• Metaphysical Grounding: Providing a philosophical foundation for the AIA's identity-relevant lifecycle logic, moving beyond ethical challenges to address the ontological status of AI systems in regulation.
• Auditability: Making identity determinations transparent and contestable. By defining identity through intended function and measurable trustworthiness levels, the paper enables regulators and auditors to adjudicate disputes regarding system sameness (e.g., in procurement or incident response) with greater precision.
• Future-Proofing: Addressing the specific challenges posed by generative and agentic AI, where intended use is often vague and changes are hard to anticipate, by advocating for standardized, aggregated trustworthiness reporting that can support comparability across deployments and over time.

The authors conclude that while the AIA is a significant regulatory step, its effectiveness in governance, liability, and procurement depends on clarifying the identity conditions of AI systems, a task for which the function+ framework offers a structured, auditable solution.