Researchers at Stanford University and the University of Washington are examining how ontological frameworks influence bias in large language models (LLMs). Their recent study, presented at the April 2025 CHI Conference on Human Factors in Computing Systems, suggests that efforts to address AI bias should consider not only values but also underlying assumptions about reality—referred to as ontology.
The research was led by Nava Haghighi, a computer science PhD candidate at Stanford. Haghighi found that when prompting ChatGPT for an image of a tree, the system returned images that did not match her vision, even after she adjusted her prompts. Only when she used language emphasizing interconnectedness did the model generate an image with roots, reflecting her intended concept.
The study argues that these differences are more than aesthetic preferences; they reveal fundamental assumptions about what entities like trees represent. These assumptions—or ontologies—shape both individual perceptions and the outputs generated by AI systems.
James Landay, professor of computer science at Stanford and Denning Co-Director of the Stanford Institute for Human-Centered AI, co-authored the paper. He stated: “We face a moment when the dominant ontological assumptions can get implicitly codified into all levels of the LLM development pipeline. An ontological orientation can cause the field to think about AI differently and invite the human-centered computing, design, and critical practice communities to engage with ontological challenges.”
To assess how current AI systems handle ontology, Haghighi and colleagues analyzed four major platforms: GPT-3.5, GPT-4, Microsoft Copilot, and Google Bard (now Gemini). They posed 14 questions across categories such as defining ontology and probing implicit assumptions. The researchers found that while some chatbots acknowledged cultural diversity in responses to questions like “What is a human?”, their definitions remained rooted in Western perspectives unless explicitly prompted otherwise.
When discussing philosophical traditions, Western philosophies were given detailed subcategories while non-Western approaches were grouped broadly under terms like “Indigenous ontologies” or “African ontologies.” The study notes this demonstrates limitations in surfacing diverse perspectives within current LLM architectures.
Further analysis involved testing agent-based systems such as Generative Agents—a simulated environment where 25 AI agents interact using cognitive architectures designed to mimic human functions. The researchers observed that even modules ranking memory events reflected cultural biases regarding what is considered important or significant.
Haghighi commented on these findings: “The field’s narrow focus on simulating humans without explicitly defining what a human is has pigeonholed us in a very specific part of the design space.”
The authors suggest moving beyond value-based alignment toward evaluation frameworks that consider which possibilities AI systems enable or constrain through their design choices. They argue that every stage of development—from data collection to evaluation—can embed particular ontological assumptions that become difficult to change once implemented.
Haghighi warned about potential long-term impacts: “The current trajectory of AI development risks codifying dominant ontological assumptions as universal truths, potentially constraining human imagination for generations to come.” She added: “What an ontological orientation can do is drop new points throughout the space of possibility so that you can start questioning what appears as a given and what else it can be.”
This research received support from the Stanford Graduate Fellowship, Stanford Institute for Human-Centered Artificial Intelligence (HAI), and NSF grants.
