AI Governance & Responsible AI

The Inquiry: The global AI governance landscape has produced 84+ ethics documents (Jobin et al. 2019), six major regulatory frameworks, and dozens of standards — all converging on the same principles (transparency, fairness, non-maleficence, responsibility, privacy). Yet responsible AI consistently fails in organizational practice (Rakova et al. 2021, Hagendorff 2020). Is this gap between principles and practice temporary (an implementation delay) or structural (a missing layer of infrastructure)? If structural, what specifically is missing?

Falsifiable formulation: If any existing AI governance framework, standard, or tool provides the operational infrastructure — not just the principles or requirements — that enables organizations to structurally enforce transparency, accountability, fairness, and human oversight as by-products of normal AI operation rather than separate compliance tasks, then the infrastructure gap claimed here does not exist.

The principles-practice gap is structural, not temporal.

The literature converges from three directions. First, the meta-analytic direction: Jobin et al. (2019) documented 84 ethics documents converging on principles while diverging on implementation; Hagendorff (2020) confirmed across 22 guidelines that operationalization consistently fails. Second, the diagnostic direction: Mittelstadt (2019) identified three missing elements (common aims, professional norms, translation methods) that medicine has and AI development lacks. Third, the practitioner direction: Rakova et al. (2021) documented that the failure is organizational — missing structures, not missing values. All three directions arrive at the same conclusion: the gap is not between intentions and effort but between principles and infrastructure.

This is the same "requirements without infrastructure" pattern S1 identified across provenance, design rationale, institutional memory, and audit compliance. S2 confirms it in a different domain: AI governance requirements are proliferating (84+ documents, six regulatory frameworks, dozens of standards) while operational implementation fails. The pattern holds across both domains, strengthening the conclusion that it is structural rather than disciplinary.


Selbst et al.'s five abstraction traps are infrastructure failures, not conceptual errors.

The most penetrating insight from the accountability literature: Selbst et al.'s (2019) five traps occur not because computer scientists misunderstand fairness but because organizations lack the infrastructure to preserve contextual richness through technical implementation. The Framing Trap (defining fairness computationally) occurs because there is no shared infrastructure for integrating social and technical perspectives. The Portability Trap (assuming transferability) occurs because governance is treated as context-free. The Formalism Trap (reducing concepts to constraints) occurs because there is no epistemic infrastructure distinguishing organizationally declared knowledge from externally validated knowledge from system-derived inferences. Reframing these as infrastructure failures rather than conceptual errors changes the solution space: not better conceptual frameworks but structural infrastructure that maintains the distinctions each trap collapses.


The assessor's regress terminates only through structural foundations.

Metcalf et al.'s (2021) "assessor's regress" — the infinite chain of justifying assessments with more assessments — is an infrastructure problem. It arises because assessment quality depends on assessor judgment without structural foundations that define what must be assessed, what counts as compliance, and what epistemic standards apply. With shared governance primitives, the regress terminates: the primitives define the assessment scope, the invariants define compliance criteria, and the epistemic classification (distinguishing types of organizational knowledge) provides the standards.

AI safety and alignment validate the same conclusion from a capability direction.

The safety tradition (Amodei et al. 2016, Russell 2019) establishes that alignment is a systems engineering problem. The alignment tradition (Christiano et al. 2017, Bai et al. 2022) demonstrates that governance can be embedded architecturally — RLHF and Constitutional AI both encode governance principles in system objectives rather than advisory guidelines. Organizations using AI systems built with these methods inherit alignment infrastructure they neither understand nor control. The governance gap is not at the model level (where alignment is advancing) but at the organizational level (where governance infrastructure for AI-augmented decisions barely exists).

Busuioc's accountability gap collapses when lineage makes authority explicit.

Busuioc (2021) documented that AI creates accountability gaps through automation bias — humans deferring responsibility to algorithmic recommendations. When AI recommends and humans rubber-stamp, the governance question of who exercised authority becomes unanswerable. The resolution is structural: every AI-augmented action must carry lineage tracing to the human authority that authorized it and the scope of that authorization. The moral buffer collapses when the lineage is explicit and structural rather than implicit and reconstructible.