The Constraints of AI in Transformational Creativity

Creativity is often characterized by qualities such as surprise, value, novelty, and individual expression. Although not as subjective as consciousness, creativity has transformed throughout history, bringing about flourishing ideas that still exist within the present day.

Understanding creativity requires a nuanced account of how ideas are conceived and how creative processes unfold.

This article is an analysis of AI’s limitations in transformational creativity and its inability to transcend predefined conceptual spaces.

Margaret Boden provides a framework that ascribes different forms of creativity: combinational, exploratory, and transformational creativity to contrast forms of creative processes and the different ways novel ideas can emerge

When discussing how an artificial agent can be creative, it certainly can fulfill these criteria, but to a partial or limited extent.

Combinational creativity is deriving a central idea based upon familiar concepts. This is often illustrated by AI, connecting multiple pieces of literature to form a seemingly new literary composition learned from different authors, recombining familiar elements that fit in an existing conceptual space. In computational terms, combinational creativity can often be implemented through algorithms that recombine stored elements according to syntactic or semantic rules.

It's important to emphasize that contemporary AI agents can embody this form of creativity, as it simply requires two converging ideas and a set of instructions. This allows agents to generate novel combinations, though not ideas that depart from the underlying conceptual framework.

Exploratory creativity explains that ideas can be created by systematically searching through conceptual spaces defined by limitations, boundaries, and existing ideals. An agent displays creativity in this sense by contriving postulates in a structured, organized manner, seeking to explore possibilities within its confined space. This is exemplified in AI by creating new novel routes by searching within a structured map defined by coordinates and traversal rules, generating new solutions within a problem space.

Once again, current AI models possess the capability of exploring conceptual spaces, locating the desired route or location dependent on a prompt. Below are instances of how AI encapsulates exploratory capabilities, thereby exemplifying creativity in defining new or valuable routes.

• Breadth-First Search (BFS): Explores the search space level by level; complete and optimal when step costs are uniform.
• Depth-First Search (DFS): Explores a single branch to its deepest point before backtracking; not guaranteed to be optimal.
• Depth-Limited Search (DLS): A depth-first strategy with a predefined cutoff to prevent unbounded exploration.
• Iterative Deepening DFS (IDDFS): Repeatedly applies depth-limited search with increasing limits, combining DFS memory efficiency with BFS completeness.
• Uniform-Cost Search (UCS): Expands nodes in order of lowest cumulative path cost, guaranteeing optimal solutions when costs are non-negative.

While these algorithms can discover optimal or unexpected solutions, they remain confined to the predefined structure of the search space itself.

Transformational creativity involves altering the structure of a conceptual space, pursuing new ideas that have not arisen before. This connotation requires an agent to traverse beyond a given constraint or boundary. Rather than merely exploring or combining ideas, it requires modifying the underlying rules, constraints, or generative principles that determine what kinds of ideas are possible in the first place.

When applying this form of creativity, a critical discrepancy arises.

Contemporary AI systems satisfy only two of the three criteria commonly associated with creativity, rendering them insufficient to be considered fully creative, under the aforementioned guidelines. While such systems may exhibit combinational or exploratory creativity within constrained frameworks, they fail to achieve transformational creativity, an essential condition for originality and the ascription of individual authorship.

When searching through a defined conceptual space, humans possess the ability to alter that dimension, reformulating its boundaries to create novel or unexpected ideas. When an artificial agent performs an exploratory or combinatory search, it references a finite conceptual space predetermined by previous inputs (statistical data); therefore, it can only be limited to generating outputs that simply represent rules defined by its inner workings. From this perspective, transformational creativity is not merely difficult for AI; it is effectively precluded by the architecture of current systems, highlighting the fundamental gap between human-like innovation and machine creativity.

The following premise-conclusion argument emerges.

P1: Transformational creativity is the ability to change rules and constraints that define a given conceptual space.

P2: Current AI systems operate by searching within a conceptual space whose rules and structures are fixed prior to the search.

P3: A system that cannot change the rules of its conceptual space cannot engage in transformational creativity.

C: Therefore, current AI systems cannot reach transformational creativity.

While this article outlines many discrepancies of AI and its ability to be creative, many techniques like meta-learning and recursive self-improvement allow agents to dynamically redefine their conceptual spaces. Meta-learning implies that the conceptual space is the learning strategies themselves, rather than the outputs given by the space. If an agent were to recursively self-improve and dynamically alter its self-modifying constraints, an AI agent would theoretically attain transformational creativity, bridging the gap between computational power and human-directed intelligence.

Such a development, however, remains speculative. Until artificial systems demonstrate the capacity to originate and revise the conceptual frameworks that govern their own learning, independent of human design, their creativity remains an extension of human intention rather than an autonomous faculty. What may emerge is not machine creativity in the human sense, but an increasingly sophisticated amplification of human-directed intelligence.