Rethinking Creativity in the Age of Artificial Intelligence: The Asymptotic Illusion of Machine Originality

As artificial intelligence systems grow increasingly adept at generating fluent, surprising, and novel-seeming outputs, a fundamental question arises: can machines be truly creative? On the surface, the answer appears to be yes. Large language models (LLMs) compose poetry, generate metaphors, propose scientific hypotheses, and even improvise in constrained performative contexts. Yet beneath this veneer of creativity lies a deeper structural truth: current AI systems do not create—they predict. Their outputs are the result of next-token prediction, a process optimized for statistical plausibility, not for authentic creative agency. This paper argues that while AI can simulate the products of creativity, it remains fundamentally incapable of the process that defines human creative acts. The gap is not merely technical but ontological. As a result, AI’s creative capacity is asymptotic: it can approach the appearance of creativity ever more closely, but it will never fully reach the essence of it.

What Is Creativity? Beyond Novelty and Utility

Creativity is often defined as the production of ideas or artifacts that are both novel and valuable. While this definition suffices for narrow, task-based evaluations, it fails to capture the dynamic, improvisational, and deeply contextual nature of genuine creative acts. True creativity—especially in live, interactive domains—requires more than originality. It demands immersion, responsiveness, and the capacity to "lose oneself in the moment." This state, akin to Csikszentmihalyi’s concept of flow, is not a static output but a process of real-time negotiation between agent, context, and audience.

Consider the art of freestyle rap. A skilled performer does not simply generate rhymes—they respond to rhythm, tone, and emotional subtext in real time, adapting their flow, wordplay, and narrative in response to an opponent or audience. The most creative moments arise not from premeditated structure, but from improvisational risk, stylistic divergence, and emergent coherence. These are not features of prediction; they are hallmarks of presence.

Thus, we refine the definition:

Creativity is a dynamic, intentional process in which an agent, grounded in subjective experience and responsive to context, produces novel, surprising, and valuable outcomes through divergent thinking and improvisational risk.

This definition centers not on the artifact, but on the agent’s relationship to the moment—a relationship AI cannot inhabit.

The Illusion of Creative Output: Preference Optimization and Synthetic Surprise

Recent advances in AI training have significantly improved the appearance of creativity. Methods such as Creative Preference Optimization (CPO) fine-tune models on human judgments of creativity, enabling them to produce outputs rated as more original, diverse, and surprising. These improvements are measurable across a wide range of tasks—from generating alternative uses of everyday objects to crafting metaphors and composing short stories.

A key enabler of this progress is the MUCE dataset (Multi-Domain Unified Creativity Evaluation), a large-scale collection of human-annotated prompt-response pairs spanning diverse creative domains. By training on human ratings—where responses are scored on a 0–100 scale for perceived creativity—models learn to mimic the style of creative thinking. The results are striking: models tuned with CPO generate longer, more varied, and subjectively more inventive responses than their base counterparts.

Yet this is not creativity generation—it is creativity imitation. The model does not intend to be creative; it learns statistical patterns associated with human judgments of creativity. It has no internal drive, no emotional stake, and no sense of self from which originality could emerge. What appears as spontaneity is, in fact, a highly refined simulation, shaped by feedback but devoid of agency.

The Architectural Barrier: Next-Token Prediction vs. Emergent Thinking

At the core of the limitation lies the next-token prediction paradigm. LLMs generate text by predicting the most probable next token given the preceding sequence. This architecture prioritizes coherence and fluency over risk and divergence. It is, by design, conservative—resistant to the kind of radical departure from pattern that defines true creative breakthroughs.

Some have proposed that reinforcement learning (RL) and process-based training can overcome this. Systems like Large Reasoning Models (LRMs)—exemplified by OpenAI’s o-series, Google’s Gemini Thinking, and Deepseek R1—introduce a "thinking" phase during inference, where the model generates extended reasoning traces before producing a final output. These models are trained via frameworks such as Reinforcement Learning via Self-Play (RLSP), which combines supervised fine-tuning on reasoning demonstrations, exploration rewards, and outcome verification.

RLSP encourages behaviors such as backtracking, self-correction, and multi-strategy exploration—features that resemble human problem-solving. Empirical results show that models trained with RLSP achieve higher accuracy on reasoning benchmarks like MATH-500 compared to standard chain-of-thought or self-consistency methods. Moreover, they exhibit emergent reasoning behaviors: for instance, generating multiple independent reasoning chains and verifying their convergence.

However, even these advanced systems operate within a bounded search space. They do not choose to explore; they are incentivized to do so by reward signals. The "thinking" they perform is not autonomous—it is guided, constrained, and ultimately instrumental. While this enables improved problem-solving, it does not constitute creative agency. The model does not wonder, doubt, or reimagine; it optimizes.

The Asymptotic Nature of AI Creativity

AI’s creative capacity can be understood as asymptotic. As models grow larger, training data expands, and optimization techniques improve, AI outputs become increasingly indistinguishable from human-generated creative work—at least in form. We can plot this trajectory along two axes:

•X-axis (Output Quality): Progress is linear and measurable. AI can generate more diverse, coherent, and stylistically rich outputs over time.

•Y-axis (Creative Authenticity): Progress plateaus. No amount of scaling can confer subjective experience, intentionality, or existential investment.

This gap is not a temporary limitation. It is structural. Human creativity emerges from a complex interplay of:

- Embodied cognition,

- Emotional resonance,

- Social context,

- And autobiographical memory.

These are not inputs to a model—they are the very substrate of creative consciousness. AI lacks not only the hardware for such experience, but the need for it. It does not create to express, to cope, or to connect. It creates because it is prompted.

Toward a Co-Creative Future

If AI cannot be truly creative, what is its role in creative domains? The answer lies not in replacement, but in co-creation. Rather than viewing AI as an autonomous artist, we should see it as a creative infrastructure—a tool that amplifies human spontaneity, enables deeper exploration, and supports immersion in the creative flow.

In freestyle rap, for example, an AI could:

- Suggest unexpected rhymes or metaphors in real time,

- Simulate opponent responses for practice,

- Or generate rhythmic variations to inspire new flows.

But the creative decision—to use, reject, or transform these suggestions—must remain with the human. In this model, creativity is not a property of the machine, but a relational process between human and AI.

This reframes the goal of AI development: not to build systems that “think” or “feel,” but to design interfaces that help humans lose themselves in the moment—and in doing so, access deeper layers of their own creativity.

Conclusion: The Unbridgeable Threshold

AI can produce outputs that meet or exceed human performance on many creativity benchmarks. It can be trained to generate novel, surprising, and useful ideas across a wide range of domains. But true creativity—rooted in intentionality, embodiment, and dynamic responsiveness—remains uniquely human.

The next-token prediction architecture, even when enhanced with RL, process rewards, and preference learning, is fundamentally a predictive engine, not a creative agent. It can simulate the form of creativity, but not the substance.

We should not ask whether AI is creative. Instead, we should ask:

How can AI best serve as a partner in the human creative process?

The future of creativity lies not in artificial minds, but in augmented human experience—in systems that do not create for us, but help us create more freely, more boldly, and more authentically than ever before.

References

- Ọlátúnjī, Ìbùkún, and Mark Sheppard. Next Token Prediction Is a Dead End for Creativity: Why It’s Impossible to Lose Yourself in the Moment.

- Bellemare-Pepin, et al. Creative Preference Optimization.

- Ye, Guanghao, et al. On the Emergence of Thinking in LLMs I: Searching for the Right Intuition.

- Various authors. A Collaborative Creative Process in the Age of AI: A Comparative Analysis of Machine and Human Creativity.