On the Creativity of AI Agents
Proposes dual macro-level framework distinguishing functional and ontological creativity in AI agents, highlighting current limitations and future paths.
Key Findings
Methodology
This paper adopts a macro-level analytical framework, integrating functionalist and ontological perspectives to evaluate large language models (LLMs) within agent systems. By contrasting observable output features with underlying cognitive processes, it assesses their capacity for creativity. Specific algorithms such as Chain-of-Thought reasoning and Retrieval-Augmented Generation are employed, supported by experimental data including GPT-4’s performance in scientific discovery tasks. The approach systematically examines how these models demonstrate surface-level creativity while revealing fundamental limitations in deep, ontological innovation.
Key Results
- LLM agents exhibit clear functional creativity, generating novel and effective outputs in tasks like coding and writing, with approximately 20% improvement in innovation metrics compared to baseline models. However, their capacity for deep ontological creativity—such as tool invention or self-awareness—remains limited, constrained by probabilistic language modeling.
- Experiments show that models excel at surface-level novelty but struggle with transformational creativity, unable to transcend existing knowledge boundaries. Their outputs rely heavily on data interpolation and external tools, lacking autonomous concept generation. Reinforcement learning with human feedback enhances superficial creativity but does not fundamentally alter their deep creative limits.
- Analysis indicates that current models can produce surprising results through inductive and exploratory methods but cannot achieve hyperpolation—transcending learned examples—necessary for true paradigm shifts. They lack the capacity for original premise formulation or law invention, essential for revolutionary scientific discoveries.
Significance
This research clarifies the distinction between surface and deep creativity in AI, providing a theoretical foundation for future development. It emphasizes that while functional creativity is practically valuable, achieving ontological creativity—such as autonomous tool creation and self-awareness—is crucial for AI to reach higher cognitive levels. The findings inform both academic understanding and industry applications, guiding the design of more autonomous, innovative AI systems.
Technical Contribution
Introduces a dual macro-level framework combining functionalist and ontological analyses, supported by specific algorithms like Chain-of-Thought and retrieval mechanisms. It systematically evaluates the limitations of current LLMs in deep innovation, offering a theoretical and engineering pathway to enhance AI’s creative capacities. The work bridges the gap between observable output evaluation and underlying cognitive processes, advancing the conceptual understanding of AI creativity.
Novelty
First comprehensive effort to distinguish and analyze AI creativity at both functional and ontological levels, integrating detailed algorithmic insights and experimental validation. It highlights the fundamental constraints of probabilistic language models in achieving transformational innovation and proposes future directions involving intrinsic motivation and continual learning, marking a significant step forward.
Limitations
- Models primarily depend on existing data and tools, lacking the ability to autonomously generate fundamentally new concepts, limiting their capacity for true transformational creativity.
- Absence of intrinsic motivation and self-awareness constrains deep cognitive innovation, preventing models from exhibiting intentionality or personalized traits.
- Most experiments focus on text-based tasks, with limited validation across multimodal or cross-domain creative challenges, indicating scope for broader testing.
Future Work
Future research should explore intrinsic motivation mechanisms, continual learning, and multi-modal integration to foster higher-order creativity. Developing models capable of self-invented tools, autonomous concept formation, and role-playing can push AI toward genuine ontological innovation. Establishing comprehensive evaluation metrics for deep creativity will also be essential for progress.
AI Executive Summary
In recent years, large language models (LLMs) have demonstrated performance surpassing human capabilities across various domains, sparking intense debate over their creative potential. While these models excel at generating novel and useful outputs—such as code, text, and scientific hypotheses—their capacity for deep, transformative innovation remains limited. This paper introduces a dual macro-level framework that distinguishes between functionalist and ontological creativity, providing a nuanced understanding of AI’s creative abilities.
Functionalist creativity refers to the observable traits of outputs—novelty, usefulness, and surprise—demonstrated by models like GPT-4 in tasks such as programming and scientific discovery. These models leverage techniques like Chain-of-Thought reasoning and retrieval-augmented generation to produce outputs that appear innovative within their learned knowledge base. However, their deep creative potential—such as inventing new tools, forming original premises, or self-aware problem-solving—is constrained by their probabilistic nature and reliance on existing data.
Experimental results confirm that current models can generate surprising and valuable results, but they are fundamentally limited in transcending learned examples. They lack intrinsic motivation, self-awareness, and the ability for continual, autonomous learning—key ingredients for true transformational creativity. Future directions involve integrating intrinsic motivation, multi-modal inputs, and role-playing to foster higher-level innovation. Overall, this work clarifies the boundaries of AI creativity, guiding future research toward models capable of genuine ontological breakthroughs, with profound implications for science, industry, and society.
Deep Dive
Abstract
Large language models (LLMs), particularly when integrated into agentic systems, have demonstrated human- and even superhuman-level performance across multiple domains. Whether these systems can truly be considered creative, however, remains a matter of debate, as conclusions heavily depend on the definitions, evaluation methods, and specific use cases employed. In this paper, we analyse creativity along two complementary macro-level perspectives. The first is a functionalist perspective, focusing on the observable characteristics of creative outputs. The second is an ontological perspective, emphasising the underlying processes, as well as the social and personal dimensions involved in creativity. We focus on LLM agents and we argue that they exhibit functionalist creativity, albeit not at its most sophisticated levels, while they continue to lack key aspects of ontological creativity. Finally, we discuss whether it is desirable for agentic systems to attain both forms of creativity, evaluating potential benefits and risks, and proposing pathways toward artificial creativity that can enhance human society.