A growing chorus of AI ethics research calls for pluralism—representing diverse values, preferences, and users. But a new paper by researchers Mushkani and Rashid, published on arXiv, argues that this framing is fundamentally incomplete. The authors contend that AI systems also impose ontologies: they define what counts as an entity, relation, feature, harm, benefit, and valid form of evidence. This hidden layer of power shapes how problems are conceptualized and solved, often before any stakeholder has a say.
The Problem of Ontological Flattening
The paper introduces the concept of ontological flattening: the conversion of situated, contested, and historically specific meanings into a restricted technical category, proxy, aggregation rule, or benchmark target that is treated as neutral and difficult to contest. For example, an AI system for urban planning might flatten complex community needs into a single metric like "average commute time," losing context about marginalized groups. The authors argue that even well-intentioned pluralistic methods can still compress categories, proxies, aggregation rules, and revision rights before affected actors have procedural standing.
The research draws on synthesis across value pluralism, pluralistic alignment, participatory and democratic AI, procedural justice, science and technology studies, and accountability research. It also incorporates aggregate themes from 11 expert interviews and three urban AI companion cases that illustrate how pluralistic approaches can improve model behavior while still committing ontological flattening.
The Pluralistic Lifecycle Governance (PLG) Framework
To address this gap, Mushkani and Rashid propose Pluralistic Lifecycle Governance (PLG)—a preliminary qualitative audit scaffold. PLG is not a validated scoring instrument; it is a framework for making the evidence and governance conditions of pluralistic AI explicit. It documents five dimensions:
| Dimension | Description |
|---|---|
| Ontological Openness | How open is the system to multiple ways of categorizing entities and relationships? |
| Epistemic Inclusion | Whose knowledge and ways of knowing are included in design and evaluation? |
| Procedural Authority | Who holds the power to define, contest, and revise categories and metrics? |
| Evaluation Pluralism | Are multiple forms of evidence and criteria used to assess system performance and harm? |
| Lifecycle Accountability | How are these commitments maintained across development, deployment, and decommissioning? |
According to the paper, PLG aims to document ontological openness, epistemic inclusion, procedural authority, evaluation pluralism, and lifecycle accountability. The authors emphasize that it is a qualitative scaffold, not a checklist, meant to surface governance conditions often hidden in technical design.
Implications for Enterprise AI Buyers
For CTOs and technology procurement leaders, this research holds a critical lesson. AI systems purchased for supply chain, logistics, or customer management come with embedded ontologies that shape what is measured, optimized, and ignored. A system that reduces worker productivity to a single metric—such as packages sorted per hour—may flatten important dimensions like safety or job satisfaction. The PLG framework offers a language for asking vendors: What categories are you imposing? Who defined them? How can they be contested?
The paper does not provide a ready-made tool, but it pushes the conversation beyond value representation toward the foundational power of categories. As enterprises accelerate AI adoption, ignoring ontological flattening risks building systems that are efficient on paper but blind to the worlds they miss.
Source: Mushkani, Rashid. "AI Pluralism and the Worlds It Misses." arXiv, June 2026.
