Ontologies in Design: How Imagining a Tree Reveals Possibilities and Assumptions in Large Language Models

Research Background and Issues

• What problems or challenges did the authors identify?
  This paper emphasizes that current analyses of Generative AI (Gen AI) and large language models (LLMs) often focus on axiology, such as issues of bias in data, while neglecting ontological considerations. The authors argue that the ontological assumptions underlying existing technologies determine the scope of what can be thought and discussed, thereby imposing limitations on future design and potentially causing long-term, implicit ontological harm.

• Why is this issue important?
  Ontology is not merely an academic inquiry into "what something is"; it determines how design shapes our modes of existence. Designs that fail to adequately consider ontology risk reinforcing existing worldviews, erasing cultural diversity, and having profound consequences for how technological systems are used and their impacts. For example, whether AI-generated outputs incorporate diverse cultural perspectives directly affects social justice and inclusivity.

• Research Motivation and Related Work
  While existing academic efforts have focused on explicit value choices and ethical compliance, this paper argues that ontology is a more fundamental yet underestimated dimension. Traditional AI research has often been criticized for adopting overly rationalist or Western-centric technological philosophies. By examining the ontological performance of generative AI and LLMs, this paper seeks to open new design paradigms that not only consider ethics but also challenge singular worldviews.

Solutions

• What methods or solutions did the authors propose?

  1. The authors propose four ontological design orientations: pluralism, groundedness, liveliness, and enactment. These orientations are used to analyze how generative AI and LLMs represent assumptions about reality or overlook diverse ontological possibilities.
  2. The authors explore these orientations through two case studies (LLM chatbots and LLM-based intelligent agent architectures) to investigate how they reveal implicit assumptions and limitations in system design.

• What is innovative about this solution?

  • Introducing ontology as a focal point for research and design fills a gap in existing studies focused on ethics and value-driven approaches.
  • The four ontological orientations provide a comprehensive analytical framework to capture deep-seated issues overlooked in generative AI design.
  • Combining ontology with design practice, the paper reveals the potential for diversified design through real-world case studies.

• What are the implementation steps and key technologies used?

  • Step 1: Define the four ontological orientations:

    • Pluralism: Investigate whether the design accommodates diverse perspectives of reality or supports a singular, dominant universal view.
    • Groundedness: Examine whether assumptions are localized and embedded in specific contexts rather than abstracted or essentialized.
    • Liveliness: Focus on whether ontological perspectives are treated as dynamically evolving rather than statically fixed.
    • Enactment: Analyze how concepts are realized in practice and the gaps between design goals and actual outcomes.

  • Step 2: Case Studies:

    1. Prompt Experiments: Pose a series of ontology-related questions to four large language models (including GPT-3.5, GPT-4, Copilot, and Bard) to observe whether their responses reflect pluralistic and dynamic ontological views and explore embedded implicit assumptions.
    2. Intelligent Agent Architecture Analysis: Study the design and implementation details of LLM-based "Generative Agents" architecture to examine the ontological assumptions about human cognition and design embedded in the cognitive architecture.

Research Findings

• What specific findings were achieved?

  1. The four ontological orientations summarize key design issues:

     • Pluralism: While LLMs can superficially display diversity in perspectives, they generally lack proactive representation of non-Western worldviews unless explicitly prompted.
     • Groundedness: LLM outputs often generalize cultural complexity through abstraction, neglecting specific contexts and cultural roots.
     • Liveliness: Although some models claim their cognitive perspectives evolve with data, they lack the ability for dynamic adjustment in practice.
     • Enactment: Evaluation processes and actual designs often fail to truly achieve theoretical goals, such as assessments based on "humanity" that may inadvertently reinforce dehumanizing perspectives.

  2. Prompt experiments reveal implicit assumptions in LLMs regarding specific questions (e.g., "What is humanity?"):

     • Models commonly define humans as "biological individuals," ignoring perspectives of interconnected existence within ecosystems.
     • Representations of non-Western perspectives often appear stereotypical or abstracted.

  3. Analysis of generative agents highlights how highly simplified cognitive models may lead to excessive simplification of human memory and emotion:

     • For instance, the importance of individual memory is programmed as static rules, neglecting the complexity of cultural, collective, or intergenerational memory.
     • Evaluation criteria (e.g., "credibility") focus on perfect simulation of human behavior, overlooking the significance of traits like "error" in defining humanity.

• What advantages does this solution have compared to existing ones?

  • This paper uniquely combines philosophical ontology with generative AI design practices as an analytical framework.
  • It expands beyond previous research that focuses solely on data bias and ethical countermeasures, uncovering implicit assumptions in technological architectures and proposing design interventions to broaden possibilities of reality.

• What were the experimental or evaluation results?

  • In prompt experiments, the study found that most language models only exhibit ontological diversity when explicitly prompted. However, such responses often appear abstract, stereotypical, and superficial.
  • Analysis of generative agent architectures revealed reductive assumptions about human cognition, with evaluation criteria (e.g., credibility measurement) potentially reinforcing these misunderstandings.

• Limitations and Future Directions

  1. Limitations:

     • The study cannot longitudinally track the long-term dynamic evolution of models and architectures (e.g., changes in ontology over time).
     • There is a lack of systematic comparison and classification standards to measure the impact of ontological design.
     • The LLMs and agent systems analyzed rely on mainstream design philosophies, limiting representation of other potential innovative architectures.

  2. Future Directions:

     • Develop new technical parameters or architectural layers to enhance ontological diversity and dynamic outputs.
     • Explore methods to better reveal ontological metaphors and assumptions in design and practice, such as through "rupture experiences" or disorientation in design.
     • Conduct deeper research into how generative AI can expand rather than constrain human imagination of reality.
     • Investigate how diverse cultural practices and traditions can be incorporated into training processes at the level of data governance.

Conclusion

This study proposes a systematic framework for analyzing the ontological gaps and corresponding harm effects in generative AI design. Through case studies, the authors clarify how the four core orientations can help uncover implicit assumptions and promote diversified, dynamic design practices. This research provides profound insights for embedding pluralistic and dynamic ontological frameworks in the future development of generative AI, offering significant originality in both philosophical depth and practical orientation within the field.