Paper ReviewInterdisciplinarySystematic Review

Who Governs AI? A Systematic Map of the Who, What, When, and How

A systematic review of 28 papers maps AI governance along four axes — who is responsible, what is controlled, when it applies, and how it works — revealing the gaps no single framework has closed.

By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.

When the European Union finalized the AI Act in 2024, it was celebrated as the world's first comprehensive AI regulation. Within months, critics pointed out that enforcement mechanisms remained vague, compliance timelines were staggered across years, and the definition of "high-risk" left vast grey zones. The EU was hardly alone. Every major jurisdiction — Washington, Beijing, New Delhi, Brasília — has produced governance documents, ethical guidelines, or regulatory proposals for artificial intelligence. The question is whether anyone has produced a coherent map of what all these frameworks actually say, where they agree, and where the gaps remain.

Batool, Zowghi, and Bano (2025) set out to draw that map.

The Research Landscape

AI governance has become one of the most crowded policy spaces in the world. Governments publish national AI strategies. Corporations release responsible AI principles. International organizations convene expert panels. Academic researchers propose framework after framework. The volume of output is extraordinary. The coherence is not.

The challenge is not a shortage of governance proposals. It is the absence of a systematic understanding of what these proposals share, where they diverge, and what they collectively leave unaddressed. Writing in AI and Ethics (Springer), Batool, Zowghi, and Bano (2025) conduct a systematic analysis of 28 peer-reviewed papers to construct a comprehensive map organized around four deceptively simple questions: WHO is responsible for AI governance? WHAT is being governed? WHEN in the AI development lifecycle does governance apply? And HOW is governance implemented?

Critical Analysis

The four-question framework structures the analysis with a clarity that the underlying literature often lacks.

<
ClaimSourceConfidenceHedge
28 peer-reviewed papers are systematically analyzedBatool et al., 2025 (abstract)HighExplicit methodological scope
The analysis maps AI governance across WHO, WHAT, WHEN, and HOW dimensionsBatool et al., 2025 (abstract)HighCore analytical framework as described
Commonalities and gaps are identified across governance frameworks from governments, corporations, and international organizationsBatool et al., 2025 (abstract)Moderate–HighThe authors analyze these three source types
The review provides a comprehensive map of AI governanceBatool et al., 2025 (abstract)Moderate"Comprehensive" is the authors' characterization; 28 papers may not capture all relevant work

The WHO dimension is perhaps the most revealing. AI governance responsibility is distributed — often ambiguously — across developers, deployers, regulators, auditors, and end users. The authors' mapping of how different frameworks assign responsibility appears to expose a fundamental coordination problem: everyone is responsible in principle, which means accountability in practice remains diffuse.

The WHAT dimension addresses the scope of governance — whether frameworks target algorithms, data, deployment contexts, outcomes, or entire sociotechnical systems. The WHEN dimension asks at which stage of the AI lifecycle governance interventions are designed to operate: design, development, testing, deployment, or post-deployment monitoring. The HOW dimension maps implementation mechanisms: regulation, standards, audits, impact assessments, certification, or voluntary codes.

By analyzing commonalities and gaps across these four axes, the authors construct something the field has lacked: a structured overview of where governance frameworks converge and where they leave blind spots.

Several critical observations warrant attention. First, the sample of 28 peer-reviewed papers, while systematically selected, necessarily excludes grey literature — the white papers, corporate guidelines, and policy briefs that constitute much of the actual governance landscape. The map may be precise but not fully representative.

Second, the four-question framework, while elegant, may underweight a fifth question that practitioners consistently raise: WHY — what theory of harm or theory of benefit motivates a given governance approach? Frameworks designed to prevent existential risk operate on fundamentally different assumptions than frameworks designed to prevent employment discrimination. Mapping them on the same axes risks false equivalence.

Third, the temporal dynamics of governance remain difficult to capture in a systematic review. Governance frameworks are not static documents; they evolve through legislative revision, judicial interpretation, and institutional learning. A map drawn from published papers captures stated intentions, not necessarily lived governance.

Open Questions

  • Enforcement gap. The HOW dimension identifies governance mechanisms, but which mechanisms actually produce compliance? Systematic mapping of mechanisms is necessary; systematic evidence on effectiveness is the harder next step.
  • Jurisdictional fragmentation. If governments, corporations, and international organizations each govern AI differently, what happens at the interfaces? Cross-border AI deployment may fall between governance frameworks rather than within any single one.
  • Power asymmetries. WHO is responsible may be less important than who has the capacity to govern. Small regulators facing large technology companies operate under structural constraints that governance frameworks may acknowledge but cannot resolve by design alone.
  • Speed mismatch. AI development cycles are measured in months. Regulatory cycles are measured in years. The WHEN dimension captures lifecycle stages, but the temporal mismatch between innovation speed and governance speed may be a critical gap of all.
  • Democratic legitimacy. Many governance frameworks are produced by technical experts or corporate actors. How — and whether — affected publics participate in governance design remains an underdeveloped question in the literature the authors survey.
  • Closing

    Batool, Zowghi, and Bano have done the field a service that is more valuable than proposing yet another framework. They have mapped the frameworks that already exist. The four-question structure — WHO, WHAT, WHEN, HOW — provides a common vocabulary for comparing governance approaches that have, until now, been described in incommensurable terms.

    What the map reveals, however, is not reassuring. The gaps the authors identify — in accountability, in lifecycle coverage, in implementation mechanisms — suggest that AI governance is still in its cartographic phase: we are learning the shape of the territory. Whether we can govern it is a question the map itself cannot answer.


    References (2)

    Batool, A., Zowghi, D. & Bano, M. (2025). [Title of article]. AI and Ethics, 5, 3265–3279.
    Batool, A., Zowghi, D., & Bano, M. (2025). AI governance: a systematic literature review. AI and Ethics, 5(3), 3265-3279.

    Explore this topic deeper

    Search 290M+ papers, detect research gaps, and find what hasn't been studied yet.

    Click to remove unwanted keywords

    Search 7 keywords →