💻 computer science

Alternative creativity: why generative AI is creative, and why it is not like us

This paper proposes "alternative creativity" as a shared mechanism of compressing vast possibility spaces into meaningful candidates via value-guided search, arguing that while generative AI and humans operate similarly in this process, they diverge fundamentally in how their underlying parameters are shaped by physical-social experience versus data archives, leading to distinct creative outputs and a new logic for human-machine collaboration.

Original authors: Jordi Linares-Pellicer, Carlos Aliaga-Torro, Isabel Ferri-Molla, Juan Izquierdo-Domenech, Joaquin Taverner

Published 2026-10-29

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Jordi Linares-Pellicer, Carlos Aliaga-Torro, Isabel Ferri-Molla, Juan Izquierdo-Domenech, Joaquin Taverner

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The question of whether a machine can be creative has long divided artists, engineers, and philosophers. On one side, some point to computer-generated images that win awards or stories that move readers, arguing that the software has finally learned to do what humans do. On the other side, critics call these systems "stochastic parrots," suggesting they merely shuffle existing words and shapes without understanding meaning, effectively laundering other people's work. Both sides, however, share a hidden assumption: that there is only one kind of creativity, the human kind, and that a machine either possesses it or lacks it entirely. This debate has reached a stalemate because it treats creativity as a single trait to be measured, rather than a complex process that might happen in different ways. To understand the new reality, we must look past the hardware and examine the actual mechanics of how both brains and computers generate ideas, and why the results they produce, while similar in some ways, stem from fundamentally different sources.

A team of researchers from the Universitat Politècnica de València has proposed a new way to view this problem, arguing that generative artificial intelligence is indeed creative, but in a manner distinct from human creativity. They call this "alternative creativity." Their analysis begins with a simple observation about the nature of art and design. Imagine a digital canvas of a fixed size. The number of possible images that could be painted on that canvas is finite, yet so vast that it exceeds the number of atoms in the observable universe. Within this unimaginably large space, almost every possible arrangement of pixels is just random noise. Only a tiny, almost invisible fraction of these arrangements would be considered a coherent picture, let alone a valuable work of art. Therefore, the act of creation is not about making something from nothing, but about navigating this massive space of possibilities to find the few points that hold value.

The researchers argue that both human brains and modern generative models solve this problem using the same two-step strategy. First, they compress the vast space of possibilities into a smaller, manageable map of what is meaningful. A human painter does not think in terms of individual pixels; they think in terms of shapes, moods, and styles, having learned to ignore the billions of useless combinations. Similarly, an artificial intelligence model learns to compress its training data into a compact internal representation, discarding the noise and keeping the patterns that make a picture look like a picture. Second, both systems navigate this compressed map. They use a form of constrained randomness to generate new candidates and then select the ones that best fit a specific goal or value. In this sense, the machine is not a passive tool but an active creator that recombines what it has learned, just as a human does.

However, the paper insists that while the process is the same, the creators are not twins. The differences lie in how they acquired their knowledge and what they can access when they create. A human creator builds their internal map through a life lived in a physical world. Their understanding of art is shaped by a body that feels hunger, pain, and joy, and by interactions with other people where actions have real consequences. A child learns from a relatively small amount of data because their brain is pre-wired by evolution to expect certain things, and because every lesson is tied to a lived experience. In contrast, a machine's map is built from a frozen archive of human testimony—trillions of words and images scraped from the internet. The machine has never felt the weight of a brush, never lost a friend, and never stood in a specific street while rain fell. It knows only what has been said and shown, not what it is like to be alive.

This difference in context leads to a crucial divergence in what each creator can produce. The machine is capable of a kind of superhuman fluency, generating photorealistic images or coherent text at speeds no human hand can match. It can explore regions of the creative space that humans have never visited, simply because it has seen more examples than any single person could. Yet, it is also limited in ways a human is not. A human can conceive of a project that no one has asked for, hold it in their mind for years, and change their own values as they live. A machine, however, cannot originate its own goals or feel the satisfaction of a job well done. It borrows its sense of value from the people who trained it. If the goal is to solve a math problem with a clear right answer, the machine can often find solutions that surpass human ability. But in the arts, where there is no single correct answer, the machine must rely on human taste to guide it, making it dependent on the very people it is meant to assist.

The researchers suggest that this dynamic changes the nature of authorship. When a human uses a machine to create, the human's role shifts from the person who makes the marks to the person who specifies the value. The human decides what is worth making, curates the results, and rejects what feels wrong. The machine provides the tireless generation of options. This does not mean the human is the sole author in the traditional sense, nor does it mean the machine is merely a tool like a paintbrush. Instead, it is a partnership where each side supplies what the other constitutionally cannot. The human provides the lived context, the intention, and the ability to judge based on feeling. The machine provides the breadth of exploration and the ability to synthesize vast amounts of information.

One consequence of this partnership is that machine-generated art tends to be more homogeneous than human art. Because the machine learns from the average of human expression, it is naturally pulled toward the middle, toward what is already familiar and popular. It struggles to reach the edges of the creative space where truly novel, strange, or difficult work lives. This is where the human advantage remains strongest. The human creator, grounded in a specific life and a specific moment, can produce work that feels authentic and connected to reality in a way the machine cannot. The machine can mimic the style of a photograph, but it cannot replicate the fact that a person stood in front of a camera at a specific time and place. This "indexicality," or the connection between the work and the real world, remains a unique human resource.

The paper concludes that the question of whether AI is creative is the wrong question. The answer is yes, but it is a different kind of yes. Generative systems are not lesser versions of human creativity, nor are they copies of it. They are a second, alternative form of creativity that traverses the same landscape of possibilities but takes different paths. They are built on a foundation of data rather than life, and they are guided by borrowed values rather than felt ones. As these systems become more common, the challenge for society will not be to determine if they are truly creative, but to understand how to collaborate with this new kind of creator. The future of art and design will likely depend on recognizing that we are no longer the only navigators in this space, and that the most valuable work will come from the unique combination of human intention and machine capability.

Technical Summary: Alternative Creativity in Generative AI

Problem Statement
The prevailing discourse regarding the creativity of generative artificial intelligence (AI) is framed as a binary question: is the machine creative or not? This dichotomy rests on an unexamined assumption that there is a single, monolithic form of "human creativity" which machines either possess or lack. The paper argues this assumption is flawed. It posits that both camps in the debate share a misconception that creativity is a unique human essence rather than a specific computational process. Consequently, the debate fails to account for the possibility that machines might be creative in a manner that is structurally similar yet constitutively different from human creativity. The paper seeks to resolve this by defining a third position: "alternative creativity."

Methodology
This article is a work of conceptual synthesis and theoretical argument, reporting no new empirical data or experiments. The methodology involves the construction and defense of a framework through the comparative integration of five distinct literatures:

Creativity Research: Specifically behavioral definitions (novelty and appropriateness) and models of generation and selection.
Cognitive Neuroscience: Leading hypotheses regarding the default mode network and frontoparietal control systems in human creative cognition.
Generative Machine Learning: The mechanics of variational autoencoders, adversarial networks, diffusion models, and large language models.
Embodied and Sociocultural Theories of Cognition: The role of the physical body, situated interaction, and collective priors in human learning.
Philosophy and Law of Art: Theories of authorship, indexicality, and the institutional definition of art.

The authors employ a "generate-and-select" loop as a unifying analytical tool, applying it to both human brains and generative models to identify shared mechanisms and constitutive divergences. Claims attributed to empirical sources are cross-referenced with primary publications, with contested evidence explicitly flagged rather than smoothed over.

Key Contributions and Theoretical Framework

1. The Shared Ground: Compression and Navigation
The paper establishes that within any fixed medium (e.g., a digital image of fixed dimensions), the space of possible configurations ( $S$ ) is finite but astronomically large, with the subset of valuable configurations ( $V$ ) being vanishingly small.

Compression: Both human brains and generative models avoid exhaustive search by learning a compressed representation (a manifold $M$ ) of the raw space. This internal model reduces the degrees of freedom from raw pixels/tokens to meaningful features (objects, styles, relations).
Navigation: Creativity is defined as a value-guided stochastic search across this compressed manifold $M$ to find points within $V$ .
The Loop: The authors argue that human creativity and machine generation both implement a "variation and selection" loop. Humans use a generative, associative component (default network) and an evaluative, selective component (control systems); machines use learned compression for candidate generation and stochastic sampling for variation, guided by a value signal (objective function or reward model).
Connectionist Principle: Both systems store competence not as an archive of specific inputs, but as adjustable connection strengths (synaptic weights in humans, model weights in AI) acquired through learning.

2. The Divergence: Four Constitutive Axes
While the abstract process is shared, the paper argues that human creators and generative models differ constitutively in four specific ways, rendering them "alternative" rather than identical or lesser:

Acquisition of the Prior: Human priors are "grown" through a lifetime of embodied, interactive experience with consequences (stakes), starting with evolutionary priors. Machine priors are "trained" via gradient descent on a frozen, tokenized archive of human testimony (the internet), lacking agency or risk.
Accessible Context: At the moment of creation, humans access a first-person, situated transaction with reality (a body, a biography, immediate relationships). Machines access a compressed archive of testimony (what has been said/shown) plus the prompt. The machine lacks the "indexical" link to a specific lived event.
Grounding of Value: Human value signals are grounded in lived experience, motivation, and sociality. Machine value signals are proxies (e.g., Reinforcement Learning from Human Feedback) that aggregate human taste but lack intrinsic grounding.
Envelope of Operations: The two systems have non-overlapping capabilities. Machines excel at hyper-realistic synthesis and high-speed variation (superhuman in specific axes) but lack the ability to originate long-horizon goals, revise their own value functions, or possess "felt" craft knowledge. Humans excel at situated intention and value revision but lack the machine's capacity for tireless, high-fidelity synthesis.

3. Derived Consequences
From this framework, the paper derives several technical and sociological consequences:

Homogenisation: Because models sample from the dense, high-probability regions of a learned manifold and are often optimized for human approval (hedonic value peaks at moderate novelty), machine output tends to cluster more tightly than human output, contracting the diversity of the cultural pool.
Limits of Attribution: The paper argues that tracing specific outputs to specific training inputs is generally indeterminate because models learn the structure (regularities) of the data, not the data itself. Memorization is a failure of compression (overfitting), not the standard mode of operation. This reframes the "theft" debate from one of copying to one of uncompensated participation in a collective learning process.
Migration of Authorship: As machines automate the "mark-making" (generation), authorship migrates to the "specification of value" (prompting, curating, constraining, and selecting). Authorship becomes a function of controlling the navigation of the manifold.
Human-Machine Collaboration: Effective collaboration is not two rivals sharing a job, but a coupling of complements. The machine provides exploratory breadth and variation; the human provides situated stakes, indexical anchoring, and the origination of goals.

Results and Claims
The paper does not present experimental results but offers a theoretical resolution to the "is AI creative?" debate.

Claim: Generative systems are creative by the behavioral standards used in creativity research (producing novel and appropriate work via a generate-and-select loop).
Claim: This creativity is "alternative." It is not a lesser version of human creativity, nor a copy, but a second instance of the process built on different constitutive foundations (training vs. living, archive vs. situation, borrowed vs. felt value).
Claim: The divergence in what is created is explained more by the difference in accessible context (lived world vs. archive) and value grounding than by the difference in substrate (neurons vs. silicon).
Claim: In domains with checkable answers (e.g., mathematics, code), machines can navigate beyond the human distribution under verifiable rewards. In the arts, where no oracle exists, machines must borrow value signals from humans, tethering them to existing distributions unless the human provides a novel value function.

Significance
The paper claims its significance lies in moving beyond the "yes/no" debate to a structural analysis of creativity. By defining "alternative creativity," it:

Legitimizes Machine Output: It acknowledges machine creativity as a real phenomenon without requiring machines to be human.
Explains Homogenisation: It provides a mechanism for why AI output tends to be similar, rooted in the statistical nature of manifold navigation.
Refines Legal and Ethical Debates: It shifts the focus of copyright and authorship debates from "retrieval of stolen data" to the nature of collective learning and the migration of authorship to value specification.
Redefines Collaboration: It offers a principled logic for human-machine co-creation based on complementary capabilities rather than competition.
Future-Proofs the Argument: The framework is presented as a finding about current constitutions, not an essentialist fence. It allows for the possibility that as AI systems move toward embodied robotics and continuous learning from their own experience, the "alternative" gap may narrow, but the distinction in value grounding remains the critical frontier.

More like this