Stanford CS547 HCI Seminar: Toward Ontological Multiplicity in AI and Computing

The Core Thesis: Moving Toward Ontological Multiplicity

Technical systems are not neutral; they encode "ontological boundaries"—assumptions about what things are and where their boundaries lie. When these boundaries are fixed and naturalized, they limit what is imaginable and designable, often reinforcing a narrow, dominant reality. Ontological multiplicity is the practice of revealing, questioning, and expanding these boundaries to create AI and computing systems that accommodate multiple, equally real ways of being.

The Danger of Embedded Ontological Assumptions

Ontological boundaries act as "cuts" (a term borrowed from physicist Karen Barad) that divide abstract phenomena into an inside and an outside. When designers and developers draw these boundaries unknowingly, they risk encoding a specific worldview into algorithms that then becomes the default reality for all users.

The "Rootless Tree" Example

To illustrate this, Nava Haghighi describes prompting an LLM to generate a picture of a tree. Despite the user's internal image of a tree including roots, the LLM consistently produced trees without roots. Even prompts mentioning the user's Iranian heritage, resulted in stereotypical imagery (desert landscapes, Persian rugs) but still no roots. Even prompts mentioning the user's Iranian heritage resulted in stereotypical imagery (desert landscapes, Persian rugs) but still no roots. Only when the prompt shifted to a philosophical assumption—"Everything in the world is connected"—did the LLM generate a tree with roots. This demonstrates how LLMs are primed toward specific ontological defaults, making alternative perspectives invisible unless explicitly articulated.

Dissolving Boundaries: The "Purple Zone"

One method for expanding ontological possibilities is to attend to "ontological glitches"—moments where a previously fixed boundary dissolves.

The Measurable Human

In affective computing, Electrodermal Activity (EDA) is used to measure stress via skin conductance. Researchers typically categorize people as "responders" or "non-responders." Those who do not meet a specific baseline are often excluded from studies, effectively rendering them invisible. This creates an ontological boundary of the "measurable human": a bounded, rational individual with objective, detectable properties.

The Purple Zone Experience

Haghighi, a non-responder, discovered a "glitch" where she suddenly became a responder during a deeply emotional meeting. This state of ontological ambiguity is termed the "Purple Zone." By building a real-time biofeedback system to notify users when they entered this zone, Haghighi observed several shifts in how participants perceived themselves:

• From Classification to Relationality: Users began to see themselves not as isolated individuals but as part of a relational unit (e.g., a "purple guest" shared with friends).
• **From Control to Cultivation: Users focused on creating the conditions (like specific activities) that might lead to the state.
• From Accuracy to Algorithmic Precision via Care: The goal shifted from accurately representing an objective property to offering an "ontological opening."

Negotiating Boundaries through User Authorship

Since not everyone experiences a glitch, Haghighi proposes "boundary negotiation" by giving users the tools to define their own categories in sensing systems.

Probing the Boundary

Using a "Wizard of Oz" approach to bypass current technical constraints, Haghighi tested two probes:

1. Event Marker: Users started from lived experience (a feeling or state) and moved toward data.
2. Pattern Finder: Users started from raw data (heart rate/energy) and moved toward meaning.

Findings in Negotiation

Participants used these tools to negotiate several types of boundaries:

• Phenomena Boundaries: A user split the category "walk" into "search walk" and "chill walk" based on their internal state.
• Subject Boundaries: A user tracking runs was actually measuring the combined entity of "human and dog."
• Signal vs. Noise: Users marked gaps in heart rate data not as sensor failure (noise), but as a way to audit the algorithm's behavior.
• Objectivity of Data: Users sought relational meanings (e.g., comparing their activity to a partner's) rather than absolute values.

A Framework for Surfacing Ontological Boundaries

To systematically identify embedded assumptions, Haghighi developed four analytical orientations:

1. Multiplicity: Asking if the system acknowledges multiple ways of being.

2. Groundedness: Examining where the assumptions are rooted.

3. Liveness: Assessing if boundaries can evolve over time.

4. Enactments: Looking at how the system actively constructs the entity it claims to measure.

Application to LLMs and Agents

Applying these orientations to current AI reveals critical gaps:

• LLM Outputs: Even when LLMs acknowledge multiple cultural definitions of "human," they almost always default to an assumption that humans are fundamentally biological individuals.
• Agent Architectingures: Systems like "Generative Agents" assume cognition is exclusively within an individual's mind.

Future Directions in AI Design

Moving forward, Haghighi suggests several shifts in the pipeline:

• Data: Using small, constrained datasets to apply "close reading" and care to understand ontological workings.
• Architecture: De-centering the individual in user models, treating relations as the foundational units rather than properties of a central person.
• Meta-Design: Investigating how AI systems that generate other systems (code/interfaces) reproduce rootless ontological assumptions and whether these can be shifted through high-level principles like "cultivation over control."