🤖 AI

AI Pluralism and the Worlds It Misses

This paper argues that AI pluralism must extend beyond mere representation of diverse values to address "ontological flattening," where AI systems impose restrictive technical categories on complex realities, and proposes a "Pluralistic Lifecycle Governance" framework to ensure epistemic inclusion and procedural authority throughout the AI development process.

Original authors: Rashid Mushkani

Published 2026-06-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Rashid Mushkani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: It's Not Just About "What" the AI Says, But "How" It Sees the World

Imagine you are trying to describe a bustling city street to a robot. You might say, "This street is a place where neighbors meet, a spot for kids to play, a risky area at night, a quiet place for reflection, and a busy commercial zone."

The paper argues that most AI systems today don't just listen to these different stories; they force the street to become one single thing so the computer can understand it. They might decide the street is simply "Safe" or "Unsafe," or "Inclusive" or "Not Inclusive."

The author calls this "Ontological Flattening."

The Metaphor: Imagine taking a rich, 3D sculpture made of clay, wood, and glass, and smashing it flat against a wall to turn it into a 2D drawing. The drawing is easier to measure and copy, but it loses the texture, the depth, and the fact that the object was made of many different materials.
The Problem: When AI "flattens" the world, it turns complex, messy human experiences (like safety, comfort, or belonging) into simple, fixed technical categories. Even if the AI gives you many different answers (pluralism), it might still be using the same flattened map to find those answers.

The Core Argument: Two Types of "Pluralism"

The paper says we need to distinguish between two things:

Outcome Pluralism (The "What"): The AI gives you many different answers.
- Analogy: A restaurant menu that offers 50 different dishes.
- The Catch: Even if there are 50 dishes, they might all be made from the exact same set of ingredients (the same "flattened" view of the world).
Procedural Pluralism (The "How"): The people affected by the AI get to decide what the ingredients are and how the menu is written.
- Analogy: The customers get to tell the chef, "Actually, 'safety' shouldn't just mean 'no crime,' it should also mean 'feeling welcome.' Let's change the recipe."

The paper argues that current AI systems often have Outcome Pluralism (lots of answers) but lack Procedural Pluralism (people can't change the rules of the game).

The Three Real-World Examples (The "Cases")

The author looked at three specific projects involving city streets and AI to show where this goes wrong:

Case A (The Art Generator): An AI was trained to generate images of "inclusive" public spaces.
- What happened: People gave feedback on what they liked. The AI learned to pick the "winning" images.
- The Flattening: The AI turned complex feelings into simple "I like this" vs. "I dislike that" votes. It erased the "I'm not sure" or "It depends" answers, forcing a binary choice where a gray area should exist.
Case B (The Street Map): An AI tried to predict which streets were "inclusive" based on photos.
- What happened: People rated streets. The AI learned to spot patterns in the photos.
- The Flattening: The AI started judging "inclusivity" based only on what it could see (like new paint or wide sidewalks). It missed the invisible stuff, like whether the local community actually felt safe or welcome there. It swapped a deep human feeling for a shallow visual clue.
Case C (The Test Score): Researchers tested AI models on how well they described city scenes.
- What happened: They compared AI answers to human answers.
- The Flattening: When humans disagreed (e.g., "Is this park safe?"), the researchers often treated that disagreement as "noise" or a mistake. But the paper argues that disagreement is actually important data! It shows that the concept is complex, not that the human is wrong.

The Solution: Pluralistic Lifecycle Governance (PLG)

The author proposes a new way to check AI systems, called Pluralistic Lifecycle Governance (PLG). Think of this not as a test that gives a score (like a report card), but as a checklist for fairness that covers the whole life of the AI project.

It asks five big questions:

Is the World Open to Change? (Ontological Openness)
- Question: Can people argue with the definitions? If the AI calls something "risky," can the community say, "No, that's actually a place of refuge"?
- Goal: The rules shouldn't be set in stone before the project starts.
Are All Voices Heard? (Epistemic Inclusion)
- Question: Does the AI only trust "expert" data, or does it also trust the lived experience of regular people?
- Goal: A grandmother's story about a street should count as evidence, not just a sensor reading.
Do People Have Real Power? (Procedural Authority)
- Question: Do people just get asked for their opinion (voice), or do they get to say "Stop" or "Change this" (power)?
- Goal: Moving from "We listened to you" to "We changed the plan because of you."
Is Disagreement Visible? (Evaluation Pluralism)
- Question: Does the report hide the fact that people disagreed, or does it show the messy reality?
- Goal: Don't average out the answers to make them look neat. Show the "I don't know" and the "It depends" answers.
Who is Responsible Later? (Lifecycle Accountability)
- Question: When the AI is launched, who can fix it if it hurts someone?
- Goal: The responsibility shouldn't end when the software is released. There needs to be a way to update the rules later.

The Takeaway

The paper concludes that AI pluralism isn't just about giving the AI more options. It's about making sure the AI doesn't force the messy, complicated, and beautiful reality of human life into a simple, rigid box.

If we want AI to be truly fair, we can't just ask it to give us 100 different answers. We have to let the people affected by the AI help decide what the questions even are, and give them the power to change the rules when the answers don't fit their world.

Technical Summary: AI Pluralism and the Worlds It Misses

Problem Statement
Current frameworks for AI pluralism primarily focus on value pluralism: representing diverse user preferences, values, or outputs. The paper argues this framing is incomplete because AI systems also impose ontologies—practical representational commitments defining what entities exist, how they relate, what constitutes evidence, and what counts as a valid output.

The core failure mode identified is ontological flattening. This occurs when situated, contested, historically specific, and embodied meanings (the "lifeworld") are converted into restricted technical categories, proxies, aggregation rules, or benchmark targets that are treated as neutral and difficult to contest. While pluralistic methods may successfully generate diverse outputs or aggregate feedback, they often operate over fixed ontologies where the definitions of categories and the rules of aggregation remain outside the authority of affected actors. This results in "outcome pluralism" (diverse answers) coexisting with "ontological monism" (a single, fixed way of defining the world).

Methodology
This paper employs a bounded conceptual and qualitative synthesis rather than a systematic review or a new controlled experiment. The methodology consists of three primary components:

Concept Mapping: A purposive theoretical sampling of literature spanning value pluralism, pluralistic alignment, participatory AI, procedural justice, science and technology studies (STS), and accountability research.
Expert Interviews: 11 semi-structured confidential interviews with experts in law, ethics, and AI. These were used for construct refinement (e.g., distinguishing procedural authority from epistemic inclusion) rather than for statistical inference or prevalence claims.
Case Analysis: Secondary analysis of three companion manuscripts focusing on urban, public-space, and vision-based systems:
- Case A: A pluralistic public-space generation dataset using Direct Preference Optimization (DPO) with community-defined criteria.
- Case B: A participatory streetscape inclusivity model using interview and focus-group data for predictive modeling.
- Case C: A reliability-aware evaluation of urban vision-language models against human annotations.

The analysis uses a five-code matrix (ontological entry point, disagreement handling, authority/recourse, proxy limitation, lifecycle continuation) to identify mechanisms of flattening across the AI lifecycle (problem framing, data generation, labeling, training, evaluation, deployment, and retirement).

Key Contributions
The paper makes three distinct contributions:

Definition of Ontological Flattening: It defines this as a specific failure mode where pluralistic inputs or outputs coexist with fixed ontologies. It distinguishes flattening from bias, misclassification, or proxy failure by asking whether the label space itself has unjustified authority and whether affected actors have standing to contest the representational commitments.
Distinction Between Outcome and Procedural Pluralism:
- Outcome Pluralism: Systems that produce, steer, or approximate diverse outputs.
- Procedural Pluralism: Systems where affected actors have standing over problem framing, categories, evidence, aggregation, evaluation, and revision.
  The paper argues that outcome pluralism is insufficient if procedural pluralism is absent.
Pluralistic Lifecycle Governance (PLG): A preliminary qualitative audit scaffold (not a validated scoring instrument) for documenting five dimensions of pluralistic AI:
- Ontological Openness: The ability to contest categories, proxies, and metrics.
- Epistemic Inclusion: The admissibility of multiple forms of knowledge (including lived experience) as evidence.
- Procedural Authority: Participation that includes decision rights (e.g., veto, pause, binding review) rather than just advisory input.
- Evaluation Pluralism: The preservation of disagreement, neutrality, abstention, and reliability data rather than their erasure.
- Lifecycle Accountability: The persistence of pluralistic commitments through deployment, monitoring, revision, and decommissioning.

Results and Empirical Findings
The analysis of the three cases illustrates how pluralistic methods can improve model behavior while still compressing categories before affected actors have procedural standing:

Case A (Generative): While community concepts were used, the DPO objective converted multi-criteria judgments into binary pairwise preferences via majority voting. Neutral comparisons (indicating ambiguity or trade-offs) were preserved but treated as a residual target rather than a valid state of the world, illustrating "procedural compression."
Case B (Predictive): Participatory labels were used to train a model, but the resulting heatmaps risked substituting visible aesthetics for situated meanings (e.g., accessibility, belonging). The governance structure lacked authority for communities to revise the map resolution or prohibit punitive uses.
Case C (Evaluation): The benchmark reported reliability and abstention, but consensus labels could still conceal unstable semantics. The study noted that dimensions closer to visible attributes were easier to measure than appraisal dimensions, and negative Krippendorff's $\alpha$ indicated systematic disagreement that should preclude treating those dimensions as stable ground truth.

The synthesis concludes that pluralism is most fragile where a lifeworld becomes a technical object. If a system begins with fixed labels, later participation can only fill in values for those labels.

Significance and Claims
The paper claims that AI pluralism must address ontological pluralism alongside value pluralism. It argues that current alignment research often aggregates heterogeneous judgments into a single training signal or treats disagreement as noise, failing to preserve contestable ways of world-making.

The significance of the work lies in shifting the focus from "how many outputs" to "who defines the ontology." It posits that legitimate abstraction is possible only when compression is bounded, documented, revisable, and proportionate, and when affected actors retain the right to contest the schema itself. The paper does not claim to provide a validated instrument for scoring AI systems; rather, it offers a framework to make the evidence and governance conditions of pluralistic AI explicit, arguing that without procedural standing over the ontology, pluralistic AI risks reproducing a monistic worldview under the guise of diversity.