Creativity Reconsidered: Generative AI and the Problem of Intentional Agency

Technical Summary: Creativity Reconsidered: Generative AI and the Problem of Intentional Agency

Problem Statement
The paper addresses the "Intentional Agency Condition" (IAC) predicament. Historically, creativity theorists (e.g., Gaut 2010; Livingston 2018) have posited that conscious intentional agency—acting in accordance with consciously represented and desired goals—is a necessary condition for creativity. This condition was largely motivated by the intuition that accidental production of novel and valuable artifacts (e.g., spilling paint to create a painting) does not constitute creativity.

However, recent advances in generative AI (specifically Large Language Models and diffusion models) present a challenge: these systems routinely produce novel, valuable, and surprising outputs that are widely judged as "creative" by both experts and laypeople, despite lacking conscious intentional agency. This creates a theoretical dilemma: either reject the IAC and accept AI creativity, or maintain the IAC and deny AI creativity (labeling it "pseudo-creativity"). The authors argue that existing theoretical responses fail to satisfactorily resolve this tension without being overly prescriptive or ignoring empirical linguistic trends.

Methodology
The paper employs a two-pronged methodological approach:

1. Critical Literature Review: The authors systematically collate and evaluate existing responses to the IAC predicament, categorizing them into "denialist" approaches (which reject AI creativity) and "accommodationist" approaches (which attempt to redefine creativity to include AI).
2. Corpus Analysis: To empirically demonstrate the shifting linguistic landscape, the authors conduct two corpus analyses:
   • Google Books Ngram Viewer: Analyzing the frequency of the bigram "creative AI" compared to "creative musician," "creative painter," and "creative engineer" from 1970 to 2022.
   • News on the Web (NOW) Corpus: Analyzing the absolute frequency of the same bigrams in web-based newspapers and magazines from 2010 to 2022.
     These analyses serve to quantify the growing tendency of Anglophone authors and journalists to predicate creativity to generative AI, particularly following the 2017 "AI boom."

Key Contributions and Proposed Solution
The authors propose a novel accommodationist framework that abandons the IAC as a necessary condition while retaining the importance of perceived intentions. Their central contribution is the redefinition of creativity as a creative ability grounded in creative functional organization.

• Creativity as Ability: Drawing on Boden's definition, the authors argue that creativity is the ability to generate novel, valuable, and surprising ideas. This ability is understood as a counterfactual capacity: an agent is creative if they are functionally organized to produce such outputs repeatedly and non-randomly.
• Creative Functional Organization: Ascriptions of creativity are inferential. Observers infer that an agent possesses a specific internal architecture capable of generating novel outputs. This inference is warranted by several signals, none of which are individually necessary but which collectively provide sufficient evidence:
  • Repeated Performance: Consistency in generating novel outputs (distinguishing ability from flukes).
  • Perceived Non-Random Contingency: Evidence of controlled unpredictability (e.g., stochastic parameters in AI) rather than pure randomness or rigid rule-following.
  • Perceived Intentions and Reflective Understanding: The presence of antecedent goals or post-hoc explanations (stronger in humans, but inferred in AI via prompt engineering).
  • Design: Knowledge that a system was engineered to produce novel outputs.
  • Responsiveness: The ability to adapt outputs in response to environmental obstacles or feedback.

Results and Analysis of Existing Responses
The paper evaluates and critiques existing theoretical positions:

• Denialist Approaches: The authors argue that "process-first" denialists (who insist on specific cognitive processes like consciousness or emotion) are overly essentialist and prescriptive. They fail to align with actual linguistic usage where "creative" is applied to AI, evolution, and nature. The "anthropomorphizing" argument (that people mistakenly attribute agency to AI) is deemed insufficient because many experts ascribe creativity to AI while explicitly acknowledging its lack of consciousness.
• Accommodationist Approaches:
  • Product-First: Rejected because it ignores the intuition that process matters (e.g., the "accidental paint" counterexample).
  • AI Satisfies IAC: Rejected as a solution to the current predicament because it relies on the unproven hypothesis of AI consciousness, which contradicts the views of many theorists who already accept AI creativity.
  • Reconceiving Intentional Agency: While the "tuning model" (Langland-Hassan 2026) offers a useful conceptual space for non-accidental but unintended creativity, the authors find it too ambiguous. It fails to adequately explain the "praise intuition" or the requirement for stable, repeated competence.

Significance and Claims
The paper claims that its "creative ability" framework successfully resolves the IAC predicament by:

1. Accommodating AI: It allows for the ascription of creativity to generative AI systems without requiring them to possess conscious intentional agency.
2. Preserving Intuition: It explains why intentional agency remains a strong heuristic for judging creativity (as it is a primary signal of functional organization) without making it a necessary condition.
3. Reconciling Praise: It reframes the "praise" associated with creativity as a general expression of approval for functional organization that yields valuable products, rather than a moral attribution of desert requiring consciousness.

The authors conclude that their approach offers a more ecumenical definition of creativity that aligns with current linguistic practices and theoretical developments, avoiding the pitfalls of both denying AI's capabilities and making speculative claims about AI consciousness. They note that further empirical research is needed to understand the extent of anthropomorphism in these judgments and to explore cross-cultural linguistic trends.