AI Pluralism and the Worlds It Misses

A new paper contends that efforts to make artificial intelligence more pluralistic overlook a deeper problem: AI systems impose their own ontologies, dictating what counts as a valid entity, relation, or harm before affected communities can weigh in [1]. The preprint, posted to arXiv on June 15, 2026, argues that the prevailing framing of AI pluralism — as a challenge of representing diverse values, preferences, or user groups — is incomplete [1]. The authors introduce the concept of “ontological flattening,” which they define as 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 [2]. The synthesis draws on value pluralism, pluralistic alignment, participatory and democratic AI, procedural justice, science and technology studies, and accountability research [2]. It also incorporates aggregate themes from 11 expert interviews and three urban AI companion cases [2]. The cases illustrate how pluralistic methods can improve or structure model behavior while still compressing categories, proxies, aggregation rules, and revision rights before affected actors have procedural standing [2]. To address this gap, the paper proposes Pluralistic Lifecycle Governance, or PLG, a preliminary qualitative audit scaffold [1]. PLG is designed to document ontological openness, epistemic inclusion, procedural authority, evaluation pluralism, and lifecycle accountability [2]. The authors stress that PLG is not a validated scoring instrument; it is a framework for making the evidence and governance conditions of pluralistic AI explicit [2]. The work arrives as large language models — a type of machine learning model trained with self-supervised learning on vast amounts of text — are increasingly deployed in high-stakes settings [6]. The paper’s emphasis on ontology echoes long-standing critiques in science and technology studies about how technical systems encode particular worldviews, often without mechanisms for contestation. By foregrounding the categories and proxies that AI systems embed, the authors aim to shift governance conversations from output diversity to the structural choices that precede model behavior.