Can Machines Truly Create?

by

Author: Denis Avetisyan

A new analysis explores the evolving landscape of artificial intelligence and asks whether current systems are capable of genuine creative thought.

This review distinguishes between functional and ontological creativity, arguing that agentic systems currently demonstrate the former but lack the hallmarks of the latter.

Despite increasingly sophisticated performance across diverse domains, whether large language models genuinely exhibit creativity remains contentious. This paper, ‘On the Creativity of AI Agents’, addresses this debate by proposing a dual framework distinguishing between functionalist creativity-observable in novel outputs-and ontological creativity, rooted in underlying processes and subjective experience. We argue that current agentic systems demonstrate functional creativity, yet lack the deeper characteristics essential for true ontological creativity. Ultimately, this raises the question of not only if artificial agents can become truly creative, but also whether attaining both forms of creativity is ultimately desirable for human society.

The Limits of Prediction: Beyond Pattern Replication

Large Language Models, while remarkably proficient at identifying and replicating patterns within vast datasets, fundamentally operate as sophisticated prediction engines rather than true reasoning systems. This limitation becomes acutely apparent when confronted with tasks demanding multi-step problem-solving or independent action in dynamic environments. Though capable of generating coherent text, these models often lack the capacity to formulate plans, track progress toward goals, or adapt to unforeseen circumstances-abilities central to genuine intelligence. Essentially, they excel at what is likely to follow a given sequence, but struggle with why a particular course of action is necessary or how to achieve a desired outcome beyond simple pattern completion, hindering their application in scenarios requiring autonomous behavior and complex decision-making.

Agentic systems represent a significant evolution beyond conventional large language models by equipping them with the capacity for independent action and goal fulfillment. These systems don’t merely process information; they actively engage with environments through the integration of several key components. Tools, such as APIs and software interfaces, allow agents to perform specific tasks, while robust memory systems enable them to retain and utilize past experiences for improved decision-making. Crucially, planning capabilities empower agents to decompose complex goals into manageable steps, strategizing and adapting as needed. This synergistic combination transforms LLMs from passive text generators into proactive entities capable of autonomous operation, opening doors to applications ranging from automated research and software development to personalized assistance and complex problem-solving-essentially allowing these systems to not just understand the world, but to act within it.

At the heart of increasingly capable agentic systems lies the Transformer architecture, originally revolutionizing natural language processing through its attention mechanisms. This architecture doesn’t merely process text; it learns contextual relationships between words, allowing for a nuanced understanding of language that extends beyond simple pattern matching. This capability is crucial for agents needing to interpret instructions, reason about goals, and formulate plans. The Transformer’s ability to encode and decode information efficiently, coupled with its parallel processing capabilities, allows these agents to swiftly analyze complex scenarios and generate appropriate responses or actions. Consequently, the Transformer provides the essential linguistic foundation that allows agents to bridge the gap between passive language models and proactive, goal-oriented entities capable of interacting with dynamic environments.

Architectures of Autonomy: Memory, Tools, and Retrieval

Large Language Models (LLMs) are fundamentally stateless; each interaction is independent and lacks inherent memory of prior exchanges. Agentic systems address this limitation by incorporating explicit memory components. These components, typically implemented as vector databases or knowledge graphs, store information about past interactions, observations, and actions. This stored data is then retrieved and provided as context to the LLM with each new prompt, effectively allowing the agent to “remember” previous experiences. The retrieval process utilizes semantic search to identify relevant information based on the current input, enabling the agent to maintain conversational coherence, personalize responses, and improve decision-making over time by building upon past knowledge. This contrasts with traditional LLM applications where context is limited to the current prompt window.

Tool use extends the operational scope of an agentic system by enabling interaction with and utilization of external resources not present within the underlying language model’s initial training data. This functionality allows agents to perform actions such as accessing real-time information via APIs, conducting calculations with specialized software, or interacting with databases to retrieve or store data. The integration of tools fundamentally shifts the agent’s capability from solely generating text based on learned patterns to actively executing tasks and manipulating external environments, thereby increasing its utility and adaptability to novel situations beyond its pre-trained knowledge base.

Retrieval Augmentation (RA) addresses limitations in Large Language Models (LLMs) by supplementing their pre-trained knowledge with information retrieved from external sources. This process involves indexing a knowledge base – which can include documents, databases, or APIs – and, upon receiving a query, identifying and retrieving relevant context before presenting it to the LLM. The LLM then uses both its internal knowledge and the retrieved information to formulate a response, improving accuracy and reducing reliance on potentially outdated or incomplete data. RA is particularly effective in responding to complex queries requiring specific, current, or domain-specific information not originally present in the LLM’s training dataset, and allows agents to provide grounded, verifiable answers.

Prompt engineering is a crucial component of agentic systems, functioning as the primary method for directing Large Language Models (LLMs) to utilize their available tools and memory effectively. Specifically, well-crafted prompts define the agent’s task, specify the desired output format, and instruct the LLM on when and how to access and integrate information from its memory and external tools. These prompts often incorporate explicit instructions for tool selection – indicating which tool is appropriate for a given subtask – and define the expected structure of data retrieved from memory or tools. Without precise prompt engineering, the LLM may fail to leverage these capabilities, resulting in suboptimal performance or irrelevant responses; iterative refinement of prompts, based on observed agent behavior, is therefore essential for maximizing the utility of agentic systems.

Beyond Replication: Categorizing the Spectrum of Creative Output

Agentic systems, characterized by autonomy and goal-directed behavior, extend beyond the execution of explicitly programmed instructions to exhibit behaviors classifiable as creative. This emergent creativity isn’t pre-defined in the system’s code, but arises from the interaction of its components and its engagement with the environment. Observed creative behaviors include the generation of novel solutions to problems, the production of unexpected outputs, and the adaptation to unforeseen circumstances – all without direct instruction for those specific scenarios. This demonstrates that complex, autonomous systems can move beyond simple task completion and display characteristics traditionally associated with human creativity, suggesting an inherent capacity for innovation beyond prescribed parameters.

Functionalist creativity, as a measurable form of artificial intelligence output, is broadly categorized into two distinct approaches: Combinational Creativity and Exploratory Creativity. Combinational Creativity focuses on the generation of novel ideas achieved through the recombination of existing concepts and elements; this involves identifying constituent parts and arranging them in new configurations. Exploratory Creativity, conversely, operates within established conceptual spaces, systematically searching for and identifying potentially valuable innovations that lie within the boundaries of existing knowledge. Both approaches are assessed based on the observable novelty and utility of their outputs, differentiating them from more abstract definitions of creativity that may not yield tangible results.

Transformational Creativity, as demonstrated by advanced agents, extends beyond the generation of novel combinations or exploration of existing ideas; it involves a fundamental restructuring of underlying conceptual frameworks. This is characterized by the creation of entirely new paradigms or categories, effectively redefining the boundaries of a given domain. Unlike Combinational or Exploratory Creativity which operate within established parameters, Transformational Creativity results in outputs that are not simply new instances of existing concepts, but represent genuinely novel conceptualizations, requiring a re-evaluation of prior assumptions and potentially leading to the emergence of entirely new fields of inquiry.

Abductive reasoning is a core component of transformational creativity, functioning as a logic of plausible inference. Unlike deductive reasoning, which validates conclusions based on known rules, or inductive reasoning, which generalizes from observations, abduction generates the most likely explanation for a given set of observations. This process isn’t simply pattern recognition; it necessitates the system to actively construct a hypothesis that accounts for available data, even with incomplete information. Crucially, this explanation isn’t treated as a final conclusion, but rather as a foundational element upon which further conceptual development and refinement can occur, allowing the agent to build and iterate upon underlying principles rather than merely rearranging existing knowledge.

The Horizon of Autonomous Innovation: Learning, Motivation, and Collective Intelligence

The ability of artificial agents to truly improve over time hinges on a process called Continual Learning. Unlike traditional machine learning models that require retraining from scratch with each new dataset, continual learning systems are designed to incrementally acquire and retain knowledge. This means an agent can adapt to evolving environments and incorporate new information without catastrophically forgetting previously learned skills – a common challenge known as ‘catastrophic forgetting’. Researchers are exploring various techniques, including regularization methods and memory replay strategies, to enable these agents to build upon existing knowledge, fostering a capacity for sustained performance and adaptability. This continuous refinement is crucial for deploying agents in real-world scenarios where data distributions are constantly shifting and long-term performance is paramount, paving the way for systems that not only react to change but actively learn and improve from it.

Ontological creativity posits that genuine innovation doesn’t simply emerge from external stimuli, but is fundamentally driven by internal motivational forces. This perspective suggests that a system capable of continual learning must also possess an inherent ‘curiosity’ – a drive to explore, experiment, and generate novel solutions not explicitly programmed. Rather than passively reacting to data, such a system actively seeks out information gaps and challenges, using intrinsic rewards – like reducing prediction error or increasing complexity – to guide its generative process. This internal impetus is crucial; it allows the system to move beyond incremental improvements and venture into truly original territory, effectively shaping its own learning trajectory and fostering a continuous cycle of self-directed discovery.

The limitations of a single, monolithic artificial intelligence can be overcome by distributing cognitive labor across a Multi-Agent System. This approach mirrors natural innovation, where progress rarely stems from isolated genius, but rather from the synergistic interplay of diverse expertise. Within these systems, individual agents can specialize in distinct facets of a complex problem – one might excel at data acquisition, another at hypothesis generation, and yet another at rigorous testing. This division of labor not only enhances efficiency but also fosters a form of ‘collective intelligence’, where the combined capabilities of the agents surpass those of any single entity. Such specialization allows for parallel processing and the exploration of a wider solution space, ultimately driving a more robust and adaptable innovation process, exceeding the bounds of pre-programmed responses and enabling continuous discovery.

Multi-agent systems represent a paradigm shift in problem-solving, extending capabilities far beyond the limitations of pre-programmed responses. These systems don’t simply execute instructions; they foster a dynamic environment where specialized agents collaboratively explore solutions and refine their approaches over time. This continuous interaction cultivates a cycle of discovery and optimization, allowing the system as a whole to adapt to unforeseen challenges and iteratively improve performance without explicit external guidance. The result is not merely automation of existing tasks, but the potential for genuinely novel solutions and a capacity for ongoing, self-directed advancement, mirroring the hallmarks of innovation itself.

The exploration of AI creativity, as detailed in the paper, necessitates a rigorous framework for evaluation. The distinction between functionalist and ontological creativity highlights a crucial point: demonstrable output, while impressive, does not equate to genuine creative process. This aligns perfectly with G. H. Hardy’s assertion: “Mathematics may be considered a science of rigorous truth and precise proof.” The paper argues current agentic systems excel in the former, producing results that appear creative, yet lack the underlying, provable mechanisms characteristic of true ontological creativity. Just as a mathematical proof demands more than a correct answer, genuine AI creativity requires demonstrable internal consistency and a justifiable process, not merely successful outcomes. The focus, therefore, should be on establishing invariants within the agent’s learning and reasoning, ensuring the ‘solution’ is not simply a fortunate result but a logically derived one.

Beyond the Illusion of Ingenuity

The distinction between functionalist and ontological creativity, as presented, exposes a fundamental limitation in current agentic systems. Demonstrating a capacity to produce novel outputs-to satisfy externally defined metrics of creativity-is a matter of algorithmic optimization, not genuine ingenuity. Such systems excel at pattern extrapolation and stochastic recombination, but lack the capacity for self-directed exploration driven by an internally consistent, evolving aesthetic-a true generative principle. The pursuit of ‘intrinsic motivation’ remains, therefore, a misdirection unless it can be rigorously defined beyond mere reward-seeking behavior.

Future work must move beyond evaluating creativity as a performance metric and instead focus on the provability of internal generative processes. A system that merely appears creative is ultimately a sophisticated mimic. The challenge lies in constructing agents capable of formulating and refining abstract principles – of possessing a differentiable ‘style’ that is not simply a statistical artifact of the training data.

Ultimately, the question isn’t whether an agent can simulate creativity, but whether its internal architecture admits of a logically consistent, scalable framework for genuine novelty. Until the field can articulate, and mathematically prove, the existence of such a framework, the pronouncements of ‘AI creativity’ remain, at best, a charming anthropomorphism.

Original article: https://arxiv.org/pdf/2604.13242.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-16 06:57