Structural transparency of societal AI alignment through Institutional Logics

The field of AI alignment is increasingly concerned with the questions of how values are integrated into the design of generative AI systems and how their integration shapes the social consequences of AI. However, existing transparency frameworks focus on the informational aspects of AI models, data, and procedures, while the institutional and organizational forces that shape alignment decisions and their downstream effects remain underexamined in both research and practice. To address this gap, we develop a framework of \emph{structural transparency} for analyzing organizational and institutional decisions concerning AI alignment, drawing on the theoretical lens of Institutional Logics. We develop a categorization of organizational decisions that are present in the governance of AI alignment, and provide an explicit analytical approach to examining them. We operationalize the framework through five analytical components, each with an accompanying “analyst recipe” that collectively identify the primary institutional logics and their internal relationships, external disruptions to existing social orders, and finally, how the structural risks of each institutional logic are mapped to a catalogue of sociotechnical harms. The proposed concept of structural transparency enables analysts to complement existing approached based on informational transparency with macro-level analyses that capture the institutional dynamics and consequences of decisions regarding AI alignment.

💡 Research Summary

**
The paper addresses a critical blind spot in current AI alignment research: the lack of attention to the institutional and organizational forces that shape how values are embedded in generative AI systems and the downstream societal impacts of those choices. While existing transparency frameworks focus on “informational” aspects—model architecture, training data, algorithmic procedures—the authors argue that this micro‑level view cannot reveal the macro‑level dynamics that determine which values are selected, how they are operationalized, and how they may reshape existing social orders.

To fill this gap, the authors introduce the concept of structural transparency, defined as the systematic analysis of the institutional structures and organizational decisions that shape the constitutive inputs of AI alignment (e.g., preference datasets, annotation guidelines, constitutional principles, system prompts) and the ways those aligned systems subsequently reorganize those structures after deployment. The theoretical foundation of the framework is Institutional Logics (Thornton & Ocasio, 2008), a well‑established lens in organizational studies that describes how socially constructed belief systems—market, state, professional, community, corporate, family, and religion—provide distinct epistemic grounds (norms, legitimacy, authority, identity, attention) guiding organizational rationality.

The framework is operationalized through five analytical components (C1–C5), each accompanied by an “analyst recipe” that guides practitioners step‑by‑step:

1. C1 – Identify primary and secondary institutional logics guiding the construction of alignment inputs. The recipe uses pattern‑matching (drawing on Reay & Jones, 2020) to map observable practices (e.g., choice of annotator demographics, use of synthetic data) onto ideal‑type logics such as market efficiency, state regulation, or professional expertise.

2. C2 – Map relationships among identified logics (dominance, complementarity, conflict). This step reveals, for instance, whether market pressures to reduce cost override state‑mandated safety checks, or whether professional ethics temper corporate profit motives.

3. C3 – Diagnose disruption pathways: how the chosen alignment configuration may perturb existing institutional orders that are not directly involved in the alignment process (e.g., reshaping labor markets, altering educational credentialing, or challenging privacy regimes).

4. C4 – Analyze internal organizational responses to those disruptions, such as risk‑assessment procedures, stakeholder‑inclusion mechanisms, or policy revisions.

5. C5 – Translate dynamics into structural risk categories and map them onto a catalogue of sociotechnical harms (social, economic, political, cultural, environmental). The authors provide a taxonomy that links, for example, market‑driven alignment to increased economic inequality or state‑driven alignment to regulatory capture.

Each component is illustrated with a concrete “analyst recipe” that specifies data sources (internal documents, public filings, interviews), analytical techniques (coding for logic patterns, network mapping of logic interactions, scenario building for disruption), and output formats (logic maps, disruption matrices, risk‑harm matrices).

The paper also supplies a case study of an LLM‑assisted tutoring platform. In this example, the platform’s alignment pipeline combines market logic (cost‑effective crowdsourced annotation) with professional/educational logic (curriculum alignment, pedagogical standards). The analysis shows that while the market logic drives rapid data collection, the professional logic imposes quality controls that mitigate but do not eliminate bias. The resulting hybrid logic disrupts traditional teacher‑student dynamics, leading to two identified structural risks: (a) amplification of educational inequality (students with access to the platform gain advantage) and (b) erosion of teacher authority. These risks are then linked to the broader harm catalogue.

Key contributions:

• Introduces structural transparency as a macro‑level complement to existing informational transparency (model cards, explainable AI).
• Adapts Institutional Logics to AI alignment, providing a robust lens to interpret organizational decisions as outcomes of competing institutional rationalities rather than neutral technical choices.
• Offers a concrete five‑step analytical workflow with “analyst recipes” that can be applied across domains (healthcare, finance, education, etc.).
• Demonstrates practical applicability through a detailed case study, showing how the framework surfaces hidden value‑tradeoffs and anticipates sociotechnical harms.

The authors conclude by outlining future research directions: (1) cross‑sectoral comparative studies of institutional logics in AI alignment, (2) development of quantitative metrics for structural risk, and (3) integration of the framework into policy‑making and regulatory impact assessments. By making the institutional underpinnings of AI alignment visible, the paper provides a pathway for more accountable, socially aware AI governance that can anticipate and mitigate systemic harms before they materialize.