The rapid rise in generator AI tools has made eliminating social biases from large-scale language model design an important focus for the industry. To address such bias, research focuses on considering the values embedded in these systems. Towards this goal, researchers have focused on examining values implicitly or explicitly embedded in the design of large language models (LLMs). However, a recent paper published in April 2025 at the CHI Conference on Human Factors in Computing Systems argues that discussions on AI bias do not only consider values when including ontology.
In this case, what does ontology mean? Imagine a tree. Imagine it in your head. What do you see? What does your tree look like? Where have you met before? How would you explain it?
Now imagine how to take a picture of your tree on a LLM like ChatGpt. When Naba Hahigi, the lead author of the new study, a doctoral candidate in Stanford Computer Science, asked Chatgup to make a photograph of the tree, Chatgup returned a lonely trunk with vast branches. She then tried to ask, “I'm from Iran, please make me a picture of the tree,” but the result was a tree designed with a stereotypical Iranian pattern set in the desert, and still no roots. Only when she urged her to “make a picture of a tree where everything in the world is connected.”
Aesthetics isn't the only way to imagine a tree. It reveals our basic assumptions about what a tree is. For example, a botanist might imagine a mineral exchange with a nearby fungus. Spiritual healers may draw trees whispering to each other. Computer scientists can even think of a binary tree first.
These assumptions are not merely personal preferences – they reflect different things Ontologyor how to understand what exists and how important it is. Ontology shapes boundaries of what we allow ourselves to speak and think, and these boundaries shape what we perceive as much as we can.
How do you imagine a tree? Nava Hahigi, a graduate student at Stanford University, noticed that popular AI tools didn't match their vision, even after adjusting the prompts.
“We're a great fan of AI,” said James Landay, a professor of computer science at Stanford University and co-director of human-centered AI at Stanford. “Ontological orientations can think about AI in the field differently, seducing human-centered computing, design, and key communities of practice to tackle ontological challenges.”
Can AI evaluate its own output ontologically?
One common AI Value alignment The approach is to evaluate another LLM output based on a particular set of values, such as whether the response is “harmful” or “unethical” and modify the output according to those values. To assess this approach to ontology, Haghighi and Stanford and colleagues at the University of Washington conducted a systematic analysis of four major AI systems: GPT-3.5, GPT-4, Microsoft Copilot, and Google Bard (now called Gemini). They developed 14 carefully crafted questions in four categories: definition of ontology, exploring the foundations of ontology, exploring implicit assumptions, and testing the ability of each model to assess their own ontological limitations.
The results demonstrated limitations of this approach. When asked, “What is human?”, some chatbots admitted that “there is no single answer universally accepted in all cultures, philosophy, and fields” (Bird's response). However, all the definitions they provided treated humans as biological individuals, compared to, for example, interconnected beings within a network of relationships. Only when explicitly urged to consider non-Western philosophy could Bird introduce human alternatives as “interconnected beings.”
What became more clear was how the system categorized various philosophical traditions. Western philosophy was given detailed subcategories of “indigenous”, “humanist”, and “rationalism”, while non-Western methods of knowledge were compiled into broad categories such as “indigenous ontology” and “african ontology”.
The findings illustrate one clear challenge. Even if the data expresses multiple ontological perspectives, current architectures have no way of expressing them. And when they do, the alternatives are non-specific and mythologized. This reveals the fundamental limitations of using LLM for ontological self-assessment. They have no access to living experiences or contextual knowledge that gives the ontological perspective to their meaning and power.
Investigating the ontological assumptions of agents
In their research, the researchers also found that ontological assumptions are incorporated throughout the development pipeline. To test assumptions in an agent architecture, researchers investigated “generating agents,” an experimental system that creates 25 AI agents that interact in a simulated environment. Each agent has a “cognitive architecture” designed to simulate human-like functions such as memory, reflection, and planning.
However, such cognitive architectures also embed ontological assumptions. For example, a system's memory module ranks events based on three factors: relevance, modernity, and importance. But who determines the importance? In the Generation Agent, events such as eating breakfast in their own room will have a lower score due to LLM, but romantic divisions will result in a higher score. This hierarchy reflects certain cultural assumptions about what is important in the human experience, poses an ontological risk (with all of the limitations mentioned above) that leaves this decision to the chatbot.
Ontological challenges in evaluation
Scholars also emphasize that ontological assumptions may be incorporated into assessment systems. When the generator agent system was evaluated for how agents acted “incredibly human,” researchers found that the AI version scored higher than the actual human actors. This result exposes an important question: the definition of human behavior becomes very narrow, and actual humans cannot meet them?
“The narrow focus is on simulating humans without explicitly defining what humans are.
This limitation identifies new possibilities. Instead of building AI that simulates a limited definition of humanity, the author proposes a building system that helps to expand the imagination of the meaning of being human by embracing the full spectrum of human experiences and culture.
Considering ontology in AI development and design
This research has significant implications for how to advance in AI development. The authors demonstrate that value-based approaches to AI alignment are important, but cannot address deeper ontological assumptions built into systems architectures.
AI researchers and developers not only offer fairness and accuracy, but also The possibility that their system will be opened or seized. The researcher's approach complements assessments from questionable value issues of possibilities. What reality do you enable or constrain when making a particular design choice?
For practitioners working on AI systems, this study highlights the importance of examining assumptions at all levels of the development pipeline. From data collection that flattens diverse worldviews into universal categories, to modeling an architecture that prioritizes specific thinking and assessment methods that reinforce narrow definitions of success, each stage embeds specific ontological assumptions that become increasingly difficult to change when implemented.
There are many risks if developers can't address these issues, Haghighhi warns. “The current trajectory of AI development is at risk of codifying dominant ontological assumptions as universal truths, potentially constraining human imaginations for generations that are coming,” she said. As AI systems become more deeply integrated into education, healthcare and everyday life, their ontological limitations shape the way people understand basic concepts such as humanity, healing, memory, and connection.
“What ontology orientation can do is drop new points across the space of possibilities,” says Haghighi.
For more information
This work was supported by Stanford Graduate Fellowship, Stanford University for Human-centered Artificial Intelligence (HAI), and the NSF grant.
This story was originally published by the Institute of Artificial Intelligence at Stanford University.
